This week, the big news is out of OpenAI. GPT-5.6 feels like, for the first time in a long time, OpenAI has come out with a better model than Anthropic. And it shipped with a new harness too.

So that’s today’s deep dive. But first, a word from our sponsor.

In Partnership with

Viktor: 5x More Work Done

Founders keep telling the Viktor team the same thing. One: "If you compare this to a VA or an employee it's not even a conversation. This is doing what an employee would do and it's doing it much much faster, better and much cheaper." Another paying team: "We can get 5x as much work done now."

Viktor is an AI employee that lives in Slack and Microsoft Teams. It connects to 3,200+ tools you already use, runs scheduled tasks on its own, and hands back real deliverables: PDFs, decks, internal apps, code PRs. You stay in control, anything sensitive waits for your approval. 40,000+ teams have already hired one.

Try Viktor free, $100 in starting credits →

Try Viktor

This Week’s AI News

Top News: We Have New Models

ChatGPT launched 5.6, which I cover in depth in today’s deep dive.

But we also saw new models from Tinker, Meta, and xAI this week.

Basically everyone besides OpenAI and Anthropic have stopped fighting over who has the smartest chatbot and started fighting over who can run your work cheapest:

• Grok 4.5 came in at $2/$6 per million tokens.

• Muse Spark 1.1 ended Meta’s open-weights era and undercut OpenAI and Anthropic by about 75% on price.

• The most interesting one might be Inkling, Mira Murati’s first from-scratch model. It’s deliberately not the strongest model out there. It’s an open base you fine-tune, the same setup Bridgewater used to build a custom Qwen that beat frontier models on domain-specific tasks.

None of these models need your attention as a consumer.

But if you’re building a product, it’s worth running the cheap ones through your evals. The price gaps are now big enough to matter.

The Other News That Mattered

• Apple is suing OpenAI and two former employees over alleged trade-secret theft for its upcoming AI hardware. Apple says candidates were told to share confidential information and even bring physical parts to interviews, and OpenAI says it has no interest in others’ secrets.

• Claude Code got an in-app browser plus a big Artifacts upgrade in the same week. Claude can now open and click through docs, designs, or your own app while it works, and Artifacts support public links and multiplayer editing, including pages made through Claude Tag in Slack. Anthropic is quietly closing the gap on the one thing OpenAI’s new app does best, which is acting on what it sees.

Resources

• Demis Hassabis wrote his framework for the AGI age, a long essay arguing AGI is probably only a few short years away and that we’re standing in the foothills of the singularity. The Nobel laureate running DeepMind is the closest thing AI has to a measured CEO voice, so the essay deserves your 15 minutes.

• OpenAI published its GPT-5.6 prompting guide, and the short version is that your old prompts are too long. Their own tests found leaner prompts scored 10 to 15 percent better on 41 to 66 percent fewer tokens. Describe the destination, stop prescribing the route.

Tools

• v0 Design Systems 2.0 lets you import your components from GitHub, npm, Figma, and Storybook so generated apps use your real design system instead of the generic default. Your prototypes finally look like your product.

Share

Deep Dive

Complete Guide: GPT-5.6

ChatGPT just had its biggest update in a year. On July 9, OpenAI shipped the GPT-5.6 models and merged ChatGPT with Codex into one app.

Combined with the leading ImageGen model, I feel like OpenAI has the best $20/month plan in AI right now. You should be taking advantage of it.

So here’s how to get the most out of ChatGPT-5.6:

1. What Actually Changed

2. The Sol, Terra, and Luna Models

3. How to Pick Your Thinking Level

4. How to Not Burn Your Limits

5. Prompt Less, Get More

Subscribe Now

1. What Actually Changed

OpenAI took two apps and made them one, with three tabs inside:

• Chat is the ChatGPT you already know.

• Codex is the coding agent, and you can keep ignoring it if you don’t code.

• The new tab is ChatGPT Work:

Work is the delegate-and-walk-away mode.

You give it a goal, it reads your files and connected apps, works for hours if it needs to, and comes back with the finished document, spreadsheet, or presentation.

Work and Codex are the same agent tuned for different jobs, which is why the two tabs look alike.

It’s on every plan, including Free, and if you read my guide to Claude Cowork, this is OpenAI’s answer, shipped to hundreds of millions of people at once.

Three more changes worth 30 seconds each:

1. OpenAI killed its Atlas browser and moved the good parts into a new ChatGPT Chrome extension, plus a browser inside the app that can visit sites, log into your accounts, and download files.

2. ChatGPT Sites turns a prompt into a live website, hosted for you, with an optional Login with ChatGPT gate for visitors. Paid plans only for now, and not yet in the EU or UK.

   

   

   

3. Computer Use got faster. The app can drive your actual computer now, opening apps and clicking buttons by looking at your screen. Hold that thought until section 7, because you should watch it happen once.

2. Sol, Terra, and Luna

GPT-5.6 comes as three models:

Sol is the senior hire. It’s like Fable. Use it for complex work and long tasks.

Terra is the mid-tier. It’s like Sonnet. It matches the old GPT-5.5 on everyday work at half the price, and it does what you asked, exactly.

Luna is the intern. It’s like Haiku.

3. Pick Your Thinking Level and Skip Ultra

Thinking levels are how long the model gets to sit with your problem before answering, from a quick glance at Light to a long stare at Max.

Then there’s Ultra, which sits in the thinking menu but behaves like a different product.

Ultra hires four copies of the model at once and splits your task between them. Every copy runs at the most expensive thinking setting, the copies can hire copies of their own, and there’s currently no dial to turn any of them down. Each one bills you separately.

Theo blew his entire 5-hour allowance in 20 minutes on a single Ultra run with fast mode on, got a manual reset, and drained that one in another 40. OpenAI has since hidden Ultra from the app’s main model slider, which tells you how they think it’s going.

Skip it.

4. How to Not Burn Your Limits

Three rules until the dust settles:

1. Stay off Ultra. You just read why.

2. Skip fast mode for now. It delivers the same answer sooner and bills you a multiple of normal usage for the privilege, and GPT-5.6 already runs marathon sessions on its own. Here's what the same tasks cost me with it off and on:

   

3. Let persistence replace thinking level. The old habit was buying quality with higher levels because models quit early. These don’t quit. Medium plus a model that finishes the job beats High on a model that needed babysitting.

5. Prompt Less, Get More

Forget the models for a minute. The most useful thing OpenAI shipped last week was a short guide on how to talk to the new ones, and its core advice fits in a sentence:

We all learned to write long, detailed prompts because the old models gave up without them. The new ones figure out the steps, so your extra instructions now get in the way.

There is one thing these models need more of, and that’s brakes. GPT-5.6 keeps going. Left alone it does the task, then the follow-up you never asked for. The fix is a stop point, written right into the ask:

Here is the complete history: [PASTE THE FULL EMAIL/SLACK THREAD(S) — do not trim].
Reconstruct: every decision made (with who made it and the message that proves it), every commitment with owner and date, decisions that were later reversed or contradicted, and open questions that everyone forgot. Chronological table. Where the record is ambiguous, say ambiguous — do not fill gaps with plausible fiction.

That’s the whole formula. The goal, then exactly where to stop and show you. Use it on anything longer than a question.

Keep in every prompt:

1. The outcome you want, stated first.

2. What “done” looks like.

3. Your hard limits, like budget, tone, and what it must never do.

4. A stop point where it pauses and shows you.

Cut from every prompt:

1. The same rule stated twice. Once is enough now.

2. Step-by-step directions for things it already knows how to do.

3. “Be concise.” It defaults shorter, so say what to keep and what to drop instead.

4. Vague personality requests like “be friendly.” Describe the behavior you want.

5. Rules that contradict each other. It wastes effort trying to obey both.

Me around the web

I joined Rupinder Singh's podcast for a full Claude Code masterclass covering how PMs should actually work with AI in 2026:

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

And today’s deep dive is the Fable 5 playbook: how to spend the free window on the work that actually deserves it, before it flips to paid credits.

But first, a word from our sponsor.

In Partnership with

Viktor: One Hire. Already Knows Your Stack.

Picture the new hire who shows up Monday already fluent in Stripe, HubSpot, Shopify, Linear, Notion, GitHub, Klaviyo, Google Ads, Meta, QuickBooks, and 3,000+ other tools.

They never ask where the forecast lives. They onboard in minutes.

Viktor is that hire. It’s an AI employee for Slack and Teams that ships real deliverables, not summaries. Marketing gets the Klaviyo audit and the campaign brief by Tuesday. Finance gets a cash flow model that updates straight from QuickBooks actuals. Ops gets the weekly manager report. Engineering gets reviewed PRs.

One install, every department benefits, every action approved by a human first. “Like adding another employee. Or two.” That’s Krysta, who runs 63 Viktor workflows at her company. 11,000+ teams now use Viktor.

Try free, $100 in credits.

Try Viktor

AI News

The Week’s Top News: Fable 5 Is Back

Anthropic just reopened access to Fable 5 after the Commerce Department lifted its export controls. The model returns under tighter filters, plus a commitment giving the U.S. government pre-release access to Anthropic’s future models.

A quick rewind on how we got here. The original order traced back to Amazon researchers who pushed past Fable’s guardrails to expose security flaws in its outputs. Anthropic argued other frontier models matched those same outputs, but Fable was the one that took the hit. The Commerce Department pulled it anyway, and for 18 days the most capable model on the market was simply unavailable to the people paying for it.

That absence mattered more than most people expected. Fable 5 had been out for barely two weeks before the ban, and people were already shipping real projects with it. Teams that had rebuilt their workflows around it got a live lesson in platform risk: your stack can lose its brain overnight, and not because of anything you did.

Here’s what the return actually looks like:

• Fable 5 is back across Claude tiers and platforms. Paid plans get it capped at half their weekly limits until July 7, then it moves to usage credits.

• The updated safety filter now blocks the specific cybersecurity issue over 99% of the time. When it triggers, you get a clear notice and a fallback answer from Opus 4.8 instead of a silent failure.

• Anthropic warned the filter could also flag harmless coding and debugging requests, though it says “the vast majority of coding work is unaffected.”

• The pre-release access deal means the U.S. government sees Anthropic’s future models before you do. That’s a first for a frontier lab, and it won’t be the last.

Here’s where I land: the model coming back is the small story. The precedent is the big one. A government pulled a frontier model off the market, extracted pre-release access as the price of return, and the lab took the deal. Every other lab watched that happen. With GPT 5.6 expected this week, we’re about to see whether OpenAI’s friendlier relationship with Washington buys it a smoother path. My bet is that “regulatory relationship” just became a line item in every AI vendor evaluation, right next to price and benchmarks.

The Other News That Mattered

• Anthropic launched Claude Science, a research app with 60+ connectable scientific databases, citation checking, and compute that runs on the lab’s own machine. That last part is the unlock: hospitals and biotechs have been blocked from AI tools by patient data rules, and Manifold Bio is already using it to rank drug candidates. Anthropic is also offering 50 grants of $30,000 each, open globally, due in two weeks.

• Anthropic also released Claude Sonnet 5, its most agentic Sonnet yet. It hits 63.2% on agentic coding versus Opus 4.8’s 69.2%, at $2 per million input tokens, and it’s now the default for free accounts. Four months ago this performance was paywalled in the most expensive model. If your product’s moat is access to expensive intelligence, that moat is a lease that keeps resetting.

• Sam Altman used an FT op-ed to call for a U.S.-led forum that sets AI safety standards and decides who can use the most advanced models, citing the IAEA as precedent. It landed just as another FT report said OpenAI discussed giving the U.S. government a 5% stake. The real question, as former White House AI advisor Dean W. Ball put it, is whether AI wealth goes directly to households or to a government that may or may not deliver.

• Palantir CEO Alex Karp criticized frontier AI labs on CNBC, saying enterprises “are paying for tokens that create no value” and the labs “are stealing weights and alpha.” Spicy, but remember he’s selling the antidote.

• Microsoft launched Frontier Company, a $2.5B unit putting 6,000 in-house engineers and sector specialists at client sites to build and run AI systems. Consulting is quietly becoming the distribution channel for enterprise AI.

Resources

• How to build a continuous eval pipeline for multi-agent systems with Gemini. This tutorial moves you from subjective testing to data-driven assessment: automated model-based grading instead of manual testing, handling unpredictable AI outputs, and wiring evaluation directly into a CI/CD pipeline with Cloud Run. If your agent testing is still “run it and eyeball the output,” start here.

• Are we ready for an agent-native memory system? Current benchmarks treat agent memory like a black box, ignoring backend data management trade-offs and operational costs. The study’s finding: no single architecture fits all scenarios. Success depends on matching the memory structure to your specific workload bottleneck. Worth reading before you commit to a memory stack.

• Google’s NotebookLM rolled out Short Video Overviews, letting you create 60-second, social-media-style educational videos from any topic or source.

Tools

• Seedance 2.5 is ByteDance’s newest frontier AI video model, generating native 30-second clips with no stitching required.

• Cursor for iOS brings Cursor to your phone. Kick off agents on the train, review the diff when you sit down.

Quick Guide: Set Up Claude Tag in Slack

Claude Tag lets you and your teammates tag @Claude like a real teammate, then have Claude work in the cloud and report back when it’s done. Here’s how to set it up:

1. Get Claude Team or Enterprise plus Slack admin permissions. In Slack, open Admin, then Apps and Workflows, search for Claude, and install it.

2. Find Claude under Apps, open it, click Home, and connect your account. Then open Claude Tag settings on the web and connect the tools it should use.

3. Run @Claude connect in Slack, copy the pairing code into Claude Tag’s browser setup, choose the channel scope, add usage credits, and launch.

4. Open Advanced in Claude Tag settings and switch from the expensive default model to Sonnet. Then run a test prompt in a dedicated Slack thread.

Pro tip: after the first run, check the admin panel for token spend, plugins used, and memory files created. You can edit or delete persistent Claude Tag memory files there.

Fundraising

• China’s Kling AI secured $2B in funding. The Kuaishou video spinoff is pushing global expansion after OpenAI shut down rival Sora. The most interesting AI video race right now isn’t happening in San Francisco.

Share

Deep Dive

The Fable 5 Playbook

Fable 5 is back, and most people are about to waste it.

Here’s the trap: you open it, run the same prompts you used on Opus 4.8, get slightly better answers, and conclude it’s not worth the fuss. That conclusion is wrong, but it’s the natural result of testing a new category of tool on the old category of task.

Fable 5 is built for work that takes a person hours, days, or weeks: full migrations, multi-day autonomous runs, problems you stopped trying to automate. Stripe pointed it at a 50-million-line Ruby migration and compressed two months of estimated team effort into a single day. That’s the category of task this model exists for.

And right now there’s a deadline attached. Through July 7, Pro, Max, Team, and select Enterprise plans get Fable 5 included for up to 50% of their weekly usage limit. After that, it moves to usage credits and the same big job costs real money. You have a short window where the most capable model available is effectively free, up to half your cap.

This deep dive is how to spend that window well. Here’s what I’m covering:

1. The one rule that decides when to use it

2. The routing config (copy-paste)

3. The big-job kickoff template (copy-paste)

4. The five prompt patterns that actually change output quality

5. Loops and the barbell strategy

6. The memory system

7. Limitations

Subscribe now

1. The one rule that decides when to use it

Weeks by hand goes to Fable 5. Minutes by hand goes to Sonnet 5.

That’s the whole routing decision. If a task would take you two weeks or more manually, it earns Fable 5 and a slice of your free window. If it’s a quick edit or a fast question, it stays on Sonnet 5, which now handles most everyday work at a fraction of the cost.

The mistake I keep seeing is people treating Fable 5 like a chat model. Asking it quick questions all day burns your included capacity on work Sonnet handles fine, and by Thursday you have nothing left for the job that actually needed the heavy machinery.

A codebase-wide refactor. A dependency migration you’ve been putting off for a quarter. A gnarly race condition nobody wants to touch. Those abandoned tasks are where Fable 5 stops looking like an upgrade and starts looking like a different product.

2. The routing config

Drop this into your CLAUDE.md so the model choice happens automatically instead of by mood:

## Model routing (through July 7)
Default: Sonnet 5 for everyday work.

Use Fable 5 only for heavy, high-payoff tasks:
- large migrations (framework, language, dependency)
- codebase-wide refactors across many files
- complex multi-step builds
- hard bugs in tangled code (race conditions, subtle state)

Rule of thumb: weeks-by-hand = Fable 5, minutes-by-hand = Sonnet 5.
Protect the Fable 5 window. Do not spend it on small edits.

# Some security-adjacent requests get rerouted to Opus 4.8
# by the new safeguards. If quality drops on one call, check for a reroute.

That last comment matters. Fable 5 came back with a new classifier that blocks one specific jailbreak technique over 99% of the time. When it triggers, your request reroutes to Opus 4.8 and you get a notice. The trade-off is false positives: legitimate coding and debugging work can occasionally get rerouted even when your intent is fine. Anthropic says the vast majority of coding work is unaffected, but if output quality suddenly drops on a single request, check whether you got rerouted before you blame your prompt.

One more thing worth knowing: Fable 5 carries a 30-day data retention policy, used to research jailbreak mitigations. Factor that in before you feed it anything sensitive.

3. The big-job kickoff template

When you point Fable 5 at a large migration or refactor, don’t say “refactor this.” Give it the whole job up front so it can hold the plan end to end. Copy and fill in:

# Task: [migration / refactor name]

## Goal
[One sentence: the end state you want.]

## Scope
- Files/dirs in scope: [paths]
- Out of scope: [paths to leave alone]

## Constraints
- Keep all tests passing.
- Preserve public APIs unless listed below.
- Match existing code style.

## Plan first
Before changing anything, map the full plan and list the files
you will touch. Wait for my go-ahead, then execute in one pass.

## Definition of done
- [ ] [check 1]
- [ ] [check 2]
- [ ] all tests green

The “plan first, then execute in one pass” line is the important part. It’s the difference between Fable 5 as a fast typist and Fable 5 as the thing that cleared a 50-million-line migration in a day. The model plans before it acts, checks its own work, and doesn’t rush to a quick output you’ll immediately have to fix.

Which means responses take longer. At high effort, a single response can take minutes. On autonomous runs, hours. The model is doing the work correctly, and correct work at this scale takes minutes. If you’re calling it through the API, extend your timeouts before you do anything else, because a timeout tuned for a ten-second Opus reply will kill a Fable run mid-plan.

4. The five prompt patterns that actually change output quality

I tested the patterns from Anthropic’s official prompting guidance against my own workflows. These five made a measurable difference. Everything else was marginal.

Pattern 1: Make it audit its own progress reports. On long runs, models can report a step as complete when it hasn’t been verified against anything real. One instruction nearly eliminates this:

Before reporting progress, audit each claim against a tool result
from this session. Only report work you can point to evidence for.
If something is not yet verified, say so explicitly. If tests fail,
say so with the output. If a step was skipped, state that.

The word “audit” is doing specific work. Softer phrasings don’t produce the same self-checking. For anything running longer than a few minutes, this goes in your system prompt, not your task prompt.

Pattern 2: Set the boundary between observing and acting. Fable 5 is noticeably more proactive than Opus 4.8. It will sometimes draft an email you didn’t ask for or create a backup branch you didn’t request. Helpful in a chat. Dangerous in an unattended pipeline. The fix:

When I am describing a problem or thinking out loud rather than
requesting a change, the deliverable is your assessment. Report
your findings and stop. Do not apply a fix until asked. Before
running anything that changes system state, confirm the evidence
supports that specific action.

Pattern 3: Give it the why, not just the what. Fable 5 uses intent context to make better judgment calls on its own. Frame requests as: “I’m building [X] for [audience]. They need [outcome]. Given that, here’s what I want you to do.” This changed my output quality more than any other single habit.

Pattern 4: Pick effort levels on purpose. High is the default and right for most real work. Xhigh is for when first-shot correctness matters more than speed, because the model reflects on and validates its own work before responding. A codebase migration warrants xhigh. A formatting pass does not. And here’s the counterintuitive part: low-effort Fable 5 often beats max-effort older models, so don’t assume you need the top dial for everything. In Claude Code, “ultrathink” in a prompt gets you max reasoning for that one turn.

Pattern 5: Rewrite your old skills instead of porting them. Skills built for weaker models tend to be prescriptive, step-by-step instructions that assume the model needs hand-holding. That same prescriptiveness can box Fable 5 in and make output worse. Go through each skill and ask: is this teaching the model something it doesn’t know, or micromanaging a model smarter than the instructions assume? Describe outcomes. Cut the steps.

5. Loops and the barbell strategy

Fable 5 was built for autonomous work, and Claude Code gives you two commands to run it that way.

/goal launches a task that runs until it’s actually done:

/goal keep researching until you can answer these 5 questions

/loop runs on an interval until you cancel it:

/loop every 30 minutes, flag any email that actually needs me

Instead of you sitting between every prompt and every fix, you design the loop once and the model handles the back-and-forth, surfacing only when it hits your success criteria.

The cost problem with long runs is real, though, and the fix is a barbell:

• Planning the loop (first 10%): Fable 5. Frontier-level planning, spec writing, loop design.

• Gruntwork and execution (middle 80%): Sonnet and Haiku subagents. Cheap, fast, repeatable.

• Final verification (last 10%): Fable 5 again. Verify against the original spec, catch errors, confirm done.

You get Fable-level judgment at the two moments it matters most, planning and verification, without paying Fable-level prices for routine execution. One run of a real research loop this way costs a few dollars in subagent tokens instead of burning your Fable window on gruntwork.

6. The memory system

None of the above hits its ceiling without persistent context. Fable 5 with a memory file is a different tool from Fable 5 without one. Anthropic’s own testing found that a persistent notes file improved Fable 5’s performance on long-horizon tasks roughly three times more than the same setup helped Opus 4.8.

The setup takes ten minutes:

1. Create a local folder like /claude-context. This is your AI brain. Put in a map of your business or project, SOPs for anything you do repeatedly, one-pagers on key projects, and a running log of decisions and outcomes.

2. Create claude-memory.md inside it, with one standing instruction: “Every time I give you major context about my business or situation, update claude-memory.md with the key details.” Now it self-updates.

3. Create claude-instructions.md with the behavior rules: how memory gets stored, output formatting, what to always and never do.

4. Connect it by pointing Claude Code at the folder via /add or referencing it in your CLAUDE.md.

For the memory file itself, use this structure:

Keep one lesson per file with a one-line summary at the top.
Record corrections and confirmed approaches alike, including why
they mattered. Do not save what the repo or chat history already
records. Update existing notes instead of duplicating. Delete
notes that turn out to be wrong.

At the end of significant sessions, close with: “Reflect on the sessions we have had. Identify core themes and lessons, and store them in the folder.” Every future session then starts with the model already knowing your world. That’s the actual mechanism behind multi-day coherence, and you own the whole thing. It lives in your local folder, portable to any model that comes next.

7. Limitations

I want to be straight about where this gets rough.

1. The reroutes are real. Security-adjacent coding work occasionally bounces to Opus 4.8 even when your intent is legitimate. You get notified, but mid-task it’s disruptive. (Workaround: check for the reroute notice before debugging your prompt, and keep the task on Opus if it keeps triggering.)

2. Latency will surprise you. Minutes per response at high effort, hours on autonomous runs. This is by design, but it makes Fable 5 wrong for anything interactive. (Workaround: fire long runs before you step away, not while you’re waiting at the keyboard.)

3. The window math punishes drift. Half your weekly limit sounds like a lot until you fritter it on medium-sized tasks. (Workaround: pick two or three genuinely heavy jobs for the week and route everything else to Sonnet, in writing, in your CLAUDE.md.)

4. One prompt gotcha: asking Fable 5 to narrate its reasoning step by step inside the visible answer can trip a classifier built to catch reasoning extraction, and you’ll get bounced to Opus on a completely benign task. If you want to see its thinking, use the structured thinking output. Don’t ask it to retype its thoughts into the reply.

Here is where I land

Everyone has access to Fable 5 now. What separates mediocre results from Stripe-level results is whether you rewrote how you talk to it.

Five-line prompts judged in ten seconds test a weaker version of this model than the one that exists. Full-context kickoffs, audit instructions, explicit boundaries, and a memory file test the real one.

You have until July 7 while it’s free. Pick the task you gave up on automating. Give it the full job. Run it at xhigh. The difference between that experience and the same prompt on Opus 4.8 is where you’ll decide for yourself whether this model is what I think it is.

That’s it for today’s deep dive.

Around the web

I joined Richa Bansal to talk about how PMs stay relevant in the AI era. We went deep on career strategy, and a few of my answers surprised even me when I heard them out loud. My five big points from the conversation:

1/ Use AI for productivity and strategy, not one or the other. Most PMs stop at automating their daily tasks. The bigger unlock is thinking strategically about where AI features belong in your product, even if you’re not on an “AI team” today. Waiting for the org to hand you an AI project is how you fall behind.

2/ Build synthetic experience if your job won’t give you the real thing. You don’t need permission to become an AI PM. Take Andrew Ng’s courses or work through Karpathy’s tutorials, then ship side projects with tools like Replit or Lovable. A shipped project beats a certificate every time with top-tier employers.

3/ Find your candidate market fit. Stop chasing generic PM roles. Niche down to the intersection of your unique experience and market demand. A specialized AI gaming PM is a far stronger position than a generalist PM, because recruiters can actually find you.

4/ Bring proof, not polish. Rehearsed talking points don’t win interviews anymore. Do your own user research, analyze the company’s competitors, and walk in with a point of view on their product’s problems before the first round. Proof of work is the new resume.

5/ Flip your weaknesses before the interviewer finds them. Job gap, industry switch, no domain experience: use AI to identify the top concerns an interviewer would have about your resume, then preemptively reframe each one. Reframe international experience as global perspective before the interviewer files it under risk.

Subscribe now

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

I pointed /goal at a research report and never once had to nudge it to keep going. It looped on its own, caught its own mistakes, checked the work against what I’d asked for, and handed back the right output. So that is today’s deep dive.

But first, a word from our sponsor + the week’s AI news.

In Partnership with

Perplexity Computer

I set up my new Mac Mini with Perplexity Computer and gave myself a challenge: build the tools I’ve always wished existed for my creator workflow. No code, no engineer. Just ask myself a question and let Computer build the answer.

One of the tools I built is a podcast guest research desk.

Every week before recording the Product Growth Podcast, I’d lose three-four hours researching the guest. So I asked myself: what do I find myself doing over and over that shouldn’t require me every time?

I described the answer out loud. Drop in a guest’s name, get back three episode angles, 15+ questions in my voice, and title options engineered against viral clip formulas. ~12K words of pre-tape intel. 45-90 seconds.

Then I said: “Computer, build me that.”

Saves me 4 hours a week. Took one prompt.

Ask yourself a question. Then tell Computer to build the answer →

Here's my build, yours to copy and rework:

Guestlab tool : Perplexity Computer

Your turn. Ask Computer to build yours:

Try here

This Week’s AI News

The Week's Top News: The Most Powerful AI on Earth Just Became Something Money Can't Buy

Anthropic launches Claude Fable 5 today, a Mythos-class model made safe for general use. It's state-of-the-art on nearly every benchmark, and the longer the task, the bigger its lead. The longer and harder the problem, the further ahead it pulls.

Anthropic shipped two versions of the same model. Fable 5 goes to the public with hard caps on cyber, bio, chemistry, and distillation. Mythos 5, the uncapped version, goes only to vetted partners: Project Glasswing cybersecurity orgs and select biology researchers, with mandatory 30-day data retention so Anthropic can watch what they do with it.

Here’s what that means in practice. A Fortune 500 with a nine-figure AI budget gets the exact same capped model as a college kid paying $20 a month. The money buys you nothing extra.

That breaks a 70-year pattern. Computing access has always scaled with money. Mainframe time, Bloomberg terminals, GPU quotas. Every tier in the history of software was a price tier, and whoever paid more got more. Anthropic just replaced price with clearance. The structure is borrowed straight from defense contracting: capability allocated by vetting, usage monitored, access revocable. Except here a private company runs the clearance system, on commercial customers, for a commercial product.

The Other News That Mattered

• Google and SpaceX AI announced compute contracts with Elon Musk’s SpaceX AI. SpaceX AI has gone from an annual ARR of $500M to $26B in just a few months, with the announcement of those two deals. No wonder its IPO popped today.

• Google DeepMind shipped Gemma 4 12B, an open multimodal model that runs on a 16GB laptop. Local, private, on-device AI keeps getting better, but people keep preferring to run frontier tokens.

• Nous Research launched Hermes Desktop, a native app for its open-source agent. The agent saves what it learns as a reusable skill, so it compounds instead of starting from zero every session.

• Meta rolled out Meta Business Agent globally across WhatsApp, Instagram, and Messenger. This is what Jensen Huang meant when he said Meta “uses AI best.” They focus on applications for their core businesses.

Tools

• Google Labs launched Dreambeans, an experiment that turns your Gmail, Calendar, and Photos into a finite daily feed of personalized stories. Pitched as the opposite of infinite scroll: it gives you 10 to 14, then stops.

• Ideogram open-sourced Ideogram 4.0 and Reve launched Reve 2.0, both betting that layout and post-generation control matter more than a better prompt. Ideogram took the top open-model spot and ships weights you can fine-tune.

Funding

• Suno raised over $400M at a $5.4B valuation, more than double its number from six months ago. Its first model built with the music industry ships in the coming months. The round closed while Universal and Sony are both still suing them, which is either confidence or a pressure test. Probably both.

Share

Deep Dive

You need to start using /goal

Here’s the thing nobody tells you about handing work to an AI agent.

There are two ways it fails you. The first is the one you know. It stops mid-task and waits. You type “keep going.” It does another turn and stops again. You type it again. An hour later you’ve said it fifteen times and you’re still the one moving the whole thing along. That's a very fast intern who needs constant supervision. You're still the one doing the real work.

The second failure is quieter and worse. The run completes. The agent says “done.” You trust it, you move on. Three days later something breaks and you find out half of it was a stub that never ran. It decided it was done. The work underneath says otherwise. You just didn’t have a way to know until it cost you.

Most people talk about /goal like it fixes the first problem. It does. The reason it matters is the second one.

That's the failure /goal is built for: the lie of self-reported "done." It catches the run that says it finished and didn't.

Here’s what I’ll cover:

1. How the loop works

2. How to draft a perfect /goal prompt

3. Practical usecases

4. Where goals break and what it costs

5. Pro tips before your first run

1. How the loop works

The pattern to use goal in Claude code, Codex is this:

1. The agent works on the task.

2. It prints the result and evidence into the conversation.

3. A second, cheaper model reads that evidence against your goal and answers one question: is the condition met?

4. If the goal is unmet, the agent takes another turn automatically.

5. If the goal is met, the run ends and control returns to you.

Two things about the checker most people miss.

First, it only reads the conversation. It can’t open files or run commands itself. The agent has to print its proof into the chat: a section list, a word count, a passing test result. If the evidence stays hidden in a file, the checker has nothing to confirm, and you’re back to taking the agent’s word for it.

Second, it’s literal. “Make this good” hands it a feeling to grade. “The document has all six sections, no placeholders, and every claim has a source link” hands it facts. Same loop, completely different result, and the only thing that changed was what you typed.

This is the difference between an agent you can leave and one you have to watch. One model works. A different model decides it’s done. They can’t be the same model, and that separation is the product.

How to set it up

In Claude Code, launch the CLI and type /goal. In Codex Desktop, go to Settings, then Configuration, and set goals = true. That’s the whole setup. While a goal is running, a live overlay shows elapsed time, turn count, and token spend. You can watch it or close the laptop.

2. How to draft a perfect /goal prompt

A working goal has three parts. Most people only write the first.

Finish line

What must be true about the final artifact? Name the specific thing and its required state. Avoid words like “good,” “polished,” or “comprehensive” unless you define what they mean in terms someone else could check.

Prove it

What evidence can the checker actually see? Ask the agent to print test results, section names, word counts, source links, or a completed checklist. If it stays hidden in a file, the checker has nothing.

Show me

What should the final handoff contain? The summary, the recommendation, the open questions, the things it couldn’t finish.

Here’s the reusable shape:

/goal [Create or complete a specific artifact]

Finish line:
- [Verifiable condition]
- [Verifiable condition]
- [Verifiable condition]

Prove it:
- [Evidence to print or test to run]

Show me:
- [Final handoff you want]

If blocked:
- [What to log, what to skip, and when to stop]

The “If blocked” section matters more than it looks. Without it, an agent burns turn after turn trying to reach a source behind a login or confirm a booking that’s sold out. Give it permission to quit cleanly and you get an honest gap instead of a runaway meter.

3. Practical use cases

Use case 1: Research reports that verify themselves

Say you need to compare five products before making a recommendation. A weak goal asks for “a comprehensive comparison.” A stronger one defines the report, the evidence, and the honest limits of the research.

/goal Build report.md comparing the five options in brief.md.

Finish line:
- Every section in brief.md is answered.
- The comparison table has no unexplained blank cells.
- Every factual claim includes a source link and source date.
- Assumptions and unresolved questions are listed separately.

Prove it:
- Print all section headings.
- Print the completed comparison table.
- Print every unresolved question and blocked source.

Show me:
- Recommend one option in three sentences.
- Name the strongest reason not to choose it.

If blocked:
- Mark the cell "Not verified" and explain why. Do not guess.

The last line protects the quality of the report. A complete-looking table with invented certainty is worse than an honest gap.

And this still needs human review. Sources can be stale, comparisons can hide important differences, and a recommendation depends on judgment the checker may not share.

Without /goal, the agent produces a table that looks complete and calls it done. With /goal, the checker runs a literal test: does every cell have a source URL? If not, it sends the agent back. You come back to something you can actually use.

The “do not guess” line is what protects the quality. A table that looks complete but invents its certainty is worse than one with an honest gap you can see.

Use /goal here when you can write a finish line a stranger could verify. A source URL, a filled cell, a passing check conditions the checker can confirm from what’s on screen. Skip it when you’re still figuring out the question. Exploratory research is a conversation, not a loop.

Use Case 2: Notes to finished draft, no placeholders

Goal loops are especially useful when the work is tedious but the output is easy to inspect.

/goal Turn notes.md into a publishable FAQ that matches example-faq.md.

Finish line:
- Every question has an answer.
- No "TODO," "[add later]," or placeholder text remains.
- Every answer is under 150 words.
- Claims not supported by notes.md are labeled for review.

Prove it:
- Print every question heading and answer word count.
- Print the results of a placeholder search.
- Print all claims labeled for review.

Show me:
- Save the finished draft.
- Summarize the three biggest editorial choices.

Use /goal here when the output has a checkable state something either passes a test or it doesn’t. Word counts, placeholder searches, section lists. Skip it when "good" is a feeling rather than a condition. If you’d know quality when you see it but can’t write a test for it, the checker has nothing to grade.

Use Case 3: Plans that run without interrupting you

You want AI to plan something with real constraints. Use /goal when the brief is settled. Put every constraint in the finish line and it runs clean, verifies what it can, flags what it can’t, and hands back something usable.

/goal Plan a 5-day Lisbon trip for two, under €1,500.

Finish line:
- itinerary.md has all five days covered.
- Every spot confirmed open on the travel dates.
- Travel between stops under 30 minutes.
- Running budget total stays under the cap.

Prove it:
- Print the day-by-day plan.
- Print the budget breakdown.
- Print every item that couldn't be confirmed.

Show me:
- The full itinerary.
- A short list of anything still needing human confirmation.

If blocked:
- Flag it. Don't quietly swap in a worse option.

The no-goal run stopped three times before it even started. It asked me for exact travel dates, whether the budget included flights, and what vibe I wanted. These are reasonable questions. They also pulled me back to my keyboard right when I wanted to walk away. That last one actually annoyed me.

The goal run answered all three of those questions itself, because the finish line already contained them. It ran clean, flagged one thing it couldn’t confirm (Sintra timed slots need advance booking it said “cannot confirm availability remotely” instead of quietly swapping in a worse option), and handed back the full plan.

Goal: 3 minutes 23 seconds. No goal: 8 minutes 2 seconds, plus my time answering questions.

The two files looked nearly identical. The difference was whether I had to stay in the room, and whether what I came back to had been checked by something other than the thing that wrote it.

Use /goal here when the brief is settled and you can put all the constraints upfront. If the goal answers every question the agent might ask, it won’t stop to ask them. Skip it when you’re still shaping what you want. A goal that needs mid-run clarification is a half-finished brief. Have the conversation first, then write the finish line.

4. Where goals break and what it costs

I’ve broken this every way there is. Every failure traces back to the same thing: the checker had nothing real to confirm.

The vague goal. “Make it production-ready” sounds decisive but defines nothing. The agent may stop early because the work looks plausible, or keep running because no condition can conclusively pass. Both burn tokens.

Better: Name the artifact, the required state, the test that proves it, and the output that confirms the test passed.

The impossible goal. The task depends on something the agent can’t access or control: a login, an approval, a sold-out reservation, a private data source. Without an escape hatch, it burns turns hammering a wall.

Better: “If blocked by something outside your control, log it and move on” lets the run end clean instead of grinding.

The bundled goal. “Research it, write it, format it, and publish it” contains several finish lines. The checker has no single condition to confirm and grades against whichever piece it noticed last.

Better: Split into sequential goals, each with one verifiable finish line. Run them in order.

What it actually costs

A goal loop runs until the condition is met. That means more turns, more elapsed time, and more tokens than a single prompt. Often a lot more. The research example above cost 4.5× more. The notes example cost 2.6× more.

Use it when the task is:

• Long enough to justify the setup

• Bounded enough to actually finish

• Verifiable with evidence the checker can see

• Valuable enough to justify the extra spend

Use a normal prompt when the task is short, exploratory, highly subjective, or waiting on human judgment.

The point is to stop supervising the work that can be checked, and save your attention for the work that genuinely can't.

5. Before your first run

Before setting a goal, answer these:

• What exact artifact should exist at the end?

• What must be true about it, in terms a stranger could confirm?

• What evidence proves those conditions, visible in the conversation?

• What should the agent do when it hits something it can’t get past?

• What still requires my judgment, and shouldn’t be in the goal at all?

Then ask the agent to critique the goal before it starts. It’ll usually find an impossible condition or a missing test before the expensive loop begins. Some of my best runs started with the model rewriting my own goal tighter than I’d written it.

A few things I wish I’d known earlier:

• Set a turn cap. 20 to 40 turns covers most real work. A slightly wrong condition loops until you kill it by hand, and you won’t be watching. Set it before you start, not after you notice it’s been running for two hours.

• Let the model write the goal. Ask Claude to draft the /goal prompt. It almost always writes a more checkable finish line than you’d write on the first try.

• Stop treating it as a coding thing. Research, notes cleanup, trip planning, audits. None of that is code. Box it into technical work and you leave most of the value sitting there.

• Constraints close the shortcuts. The agent takes the shortest path to your condition, which is sometimes the path you didn’t want. “Do not guess. Do not use placeholders. Do not quietly skip items.” These lines go into almost every goal I write now.

The command takes ten seconds to learn. The thing that takes longer is writing down what finished means before you walk away, in words a checker can confirm without reading your mind.

Build that habit and you stop being the thing that keeps the work moving. You come back to work that was checked by something other than the thing that did it.

Everyone else is still in the chair, typing “keep going,” or trusting a “done” they shouldn’t.

Subscribe now

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

Three weeks running it, I stopped re-deriving conclusions I had already reached. So that is today’s deep dive.

But first, a word from our sponsor + the week’s AI news.

I gave one of my recurring research jobs to an agent

I built a Podcast Guest Prep agent on Hyperagent.

First run, it shipped me a briefing on Howie Liu: recent appearances, strongest public arguments, sharp question angles, and what not to ask.

Then I saved it as an agent.

Now every time I have a new guest, I point that same agent at a name and it ships back a sharper prep doc because it has learned my taste and recurring question framings.

That is one use case. Founders and operators are using Hyperagent to build whole fleets of agents that run real parts of the business.

Built by the Airtable team. Separate product.

Hyperagent is offering $1,000 in inference credits to 1,000 of our subscribers. Claim yours here.

Try Hyperagent

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week's Top News: xAI Finally Has a Real Answer to Claude Code

Elon publicly admitted xAI had fallen behind on coding. Then they shipped.

Grok Build is xAI’s terminal-based coding agent, now in early beta for SuperGrok and X Premium Plus subscribers. The headline feature is plan mode: before it touches a single file, it shows you every proposed change and waits for your sign-off. That one decision tells you a lot about what they learned watching Claude Code ship.

It also runs up to eight sub-agents in parallel, each working its own branch of your codebase. You stop waiting for one pass to finish before the next one starts. The context window is 2 million tokens.

If you manage engineers or work closely with a technical team, the coding agent market just got meaningfully more competitive. Claude Code, Codex, and Grok Build are now three real options with different pricing models and different strengths. Your team will start asking which one they should standardize on. The answer matters because these tools now generate 70% of committed code at companies like Uber, and the ROI question is becoming harder to avoid.

The plan mode is the feature worth paying attention to. It forces the agent to explain its reasoning before acting. That is not just good engineering practice. It is the difference between a tool your team trusts and one they stop using after one bad deploy.

If you want a breakdown of how PMs are actually using coding agents to ship faster, Product Growth covers that in depth.

The Other News That Mattered

• Anthropic released Opus-4.8, announced it is at $47B ARR, and has raised another round at a $965B valuation. It is now worth more than OpenAI. Interestingly, Opus-4.8 doesn’t beat GPT-5.5 on agentic terminal coding, as Anthropic continues to withhold Mythos-class models from us.

• Andrej Karpathy joined Anthropic to work on pre-training, the compute-intensive phase that shapes Claude’s core capabilities. An OpenAI co-founder chose Anthropic over returning to OpenAI. That sequence is the signal.

• Uber burned its entire 2026 AI budget in four months and the COO isn’t sure it was worth it. 95% of engineers use AI tools monthly, 70% of committed code is AI-generated, and the link to shipping more consumer features still isn’t there. This is the conversation every company will have in the second half of 2026.

• OpenAI pitched GPT-5.5 Cyber to Japan for its government and private sector, covering 15 critical infrastructure sectors. The model is built specifically for cyber defense, with lower guardrails for verified defenders. AI is now a formal piece of national security deals.

• Anthropic co-founder Chris Olah spoke at the Vatican alongside Pope Leo XIV’s release of his AI encyclical. Olah said frontier AI labs operate inside incentives that can conflict with doing the right thing, and the world needs critics from the outside. A striking thing to say at the Vatican. Also completely true.

Resources

• Codex now supports /goal for long-running tasks, plus Computer Use for navigating Mac apps. If you’ve been waiting for it to handle actual multi-step work, this is the update to try.

• Qwen 3.7 Max is Alibaba’s agent model built for long autonomous sessions. Their headline test: a 35-hour GPU optimization run with over a thousand tool calls. Take the numbers with some skepticism, but the direction is real.

Tools

• PollyReach gives your AI agents their own phone numbers to make and receive real calls for lead qualification, support, and bookings. The gap between chat AI and phone AI just got smaller.

• Chert (YC-backed) lets you build and deploy AI directly on iMessage. The insight: the best interface for most people is the one already open on their phone.

Funding

• SpaceX is preparing the largest IPO in history at a $1.75 trillion valuation, and Google just opened talks to host data centers in orbit. Google already owns 6.1% of SpaceX, a stake now worth roughly $107 billion on paper. Anthropic, xAI, and Google are all in talks for orbital compute capacity. SpaceX goes public this summer with its biggest customers already locked in as shareholders.

Share

Your AI Forgets Everything Between Conversations. Gbrain fixes this.

GBrain is open source, free, and built by the CEO of Y Combinator. Three weeks running it changed how I think about what AI is actually for.

Garry Tan has been running Y Combinator since 2022. He has also been building his own AI agent for most of that time. The problem he kept running into was that every session started from zero. Context he had built over weeks was gone. People he had introduced it to were strangers again. Decisions he had worked through were unrecognised.

He spent his nights building a fix. He called it GBrain.

Most people hit this wall and assume it is just how AI works. You explain your situation. You get a good answer. You come back the next day and explain it again from scratch. You are the only one with memory in the relationship. The AI is always meeting you for the first time.

GBrain changes the structure. It builds a persistent brain on top of your agent (Hermes or OpenClaw) from everything you talk about, save, and think through. Every session adds to it. The thing you mentioned three months ago is searchable today.

Last month I was sitting on a complicated situation I had been circling for weeks. Notes in different apps. Articles saved but not connected. Thoughts I had written down and promptly lost. I asked Hermes to pull together what I actually knew about it. In under a minute it returned a clear synthesis with sources going back six weeks. I had been using it as a sounding board. The whole time it had been keeping score.

That is the shift GBrain makes. Here is how to build it.

Today’s Post

1. What GBrain Actually Is

2. One Install. Then It Grows Every Night.

3. Three Things That Change When It Runs

4. The Ideas Your Notes Are Hiding From You

5. Here Is Where I Land

1. What GBrain Actually Is

GBrain is a memory system that builds itself from how you already work.

Every time you talk to Hermes, it listens. You mention someone you know. Their page starts building in the brain. You share an article. It gets read, summarised, and filed. You describe a decision you are sitting on. It gets stored in your exact words. You record a call. It gets transcribed, broken into people and topics, and added to every relevant page.

At night it runs a maintenance cycle on everything that came in. It updates what it thinks it knows when new information contradicts the old. It flags when two things you saved say different things about the same topic. By morning the pages are current.

When you ask it something, it searches those pages. Your pages. The ones built from your actual conversations and reading. Then it writes you an answer with a citation back to the exact source it drew from. If it does not have the information, it says so. It never guesses.

Every AI tool you have used knows the internet. GBrain knows your life.

I have been covering how this category evolves. Karpathy’s autoresearch was the first version that felt genuinely different. Then his wiki pattern. Karpathy’s system stored what you put in. GBrain builds its own understanding of what you put in. That is a different thing.

2. One Install. Then It Grows Every Night.

GBrain runs inside Hermes Agent or OpenClaw on Telegram. It does not work standalone. If you have not set up Hermes yet, start there first: [link to Hermes setup].

Once Hermes is running:

bun install -g github:garrytan/gbrain
gbrain init --pglite
gbrain skillpack scaffold --all
gbrain doctor

gbrain doctor shows a health score out of 100. A healthy brain scores 70 or above.

Install the night cycle before anything else:

gbrain autopilot --install

Without this, the brain stops growing after you close Hermes. The autopilot schedules an overnight maintenance run that updates pages, repairs citations, and flags contradictions between sources. This is what makes it compound.

gbrain integrations

Find the meeting-sync recipe. Every call you have from that point gets added to the brain automatically. Every person gets a page. Every topic discussed gets a timeline entry.

The brain starts empty. From the first message you send to Hermes, the signal detector starts filing. You mention someone. Their page starts building. You describe a decision. It gets filed in your exact words. You share a link. It becomes a brain page with analysis. You do not have to do anything deliberately. Just talk to Hermes normally.

By week two the brain has enough density to start returning useful answers. The outputs in this post reflect three weeks of normal use.

3. Three Things That Change When It Runs

The prompts below are the exact messages to send Hermes on Telegram. Run them from week two onward.

Use Case 1: The Shopping Agent That Already Knows You

Every time you shop for something you explain your preferences from scratch. GBrain builds a preference profile from your natural conversations. You mention you loved a jacket. Filed. You say you are done with fast fashion. Filed. You describe the running shoe that destroyed your knees. Filed. Three months later you ask what you already know.

Prompt:

“I’m looking for a new everyday bag. Based on everything I’ve said about how I work and travel, what should I actually be looking for?”

What comes back:

Every constraint you have mentioned across every conversation: the things that have broken, the sizes that have not worked, the material you said you were done with, the feature you said you wanted to try next time. Synthesised and cited.

You are not describing your preferences again. You described them once. The brain kept them.

More prompts to run:

• “What have I said I hate about bags I have owned?”

• “I loved the last [jacket / shoes / bag] I bought. What did I say made it work?”

• “What did I rule out last time I was shopping for [item] and why?”

Use Case 2: When You Have Been Reading About Something for Weeks and Cannot Synthesize It

You have been following the AI agent space for two months. Demos, threads, papers, conversations with people building with it. The signal is real but it is scattered. You have tabs open from six weeks ago. Notes that do not connect to each other. No clear picture.

The brain has been building one the whole time.

Prompt:

“I’ve been tracking AI agents for two months. Pull together everything I’ve saved and written about it. What do I actually know?”

What comes back:

A synthesis across every page the brain built on the topic. What the demos showed. What the people you spoke to believed. What your own notes concluded. Where the sources contradict each other. All cited.

You already did the research. The brain just never let you lose it.

More prompts to run:

• “What’s the gap between what I’ve been reading about AI agents and what I actually believe?”

• “What keeps coming up every time I research [a new space I’m following]?”

• “What did I conclude and then stop acting on?”

Use case 3: The Conclusion You Keep Losing and Re-Deriving

Three years ago you figured out you do your best thinking before 9am with no meetings. It took you two years to get there. You have since let it erode. Every six months you rediscover it and start protecting the mornings again. Then it erodes again.

That conclusion lived in three different notebooks, a deleted app, and a journal you no longer use. It was never searchable. You keep re-deriving the work you already did.

GBrain holds it. The reasoning, the evidence, the date you figured it out.

Prompt:

“What have I figured out about how I work best? I know I’ve mentioned this before.”

What comes back:

Every insight you have mentioned about your own patterns, pulled from across every conversation. Ranked by how often the same conclusion appears in different places. The brain returns the evidence that built it, not just the conclusion.

The insight you keep rediscovering becomes something you only had to discover once.

More prompts to run:

• “What patterns have I noticed about the decisions I end up regretting?”

• “What did I figure out about my energy levels that I keep ignoring?”

• “What’s something I told myself I would do differently that I still haven’t?”

4. The Ideas Your Notes Are Hiding From You

Think of your brain as a library of everything you have ever read, saved, and thought. Brainstorm mode takes two ideas from that library that are close to your question and puts them together. It finds the connection between a note you wrote in February and an article you saved in October that you never put in the same sentence. The brain had.

LSD mode, Lateral Synaptic Drift, pulls from the far edges of your library. Things that have nothing obvious to do with your question. It forces a collision. Every idea must invert at least one thing you have taken for granted. The judge in the source code rejects anything too obvious.

I ran brainstorm before making a decision I had been circling for two months. It returned a connection between something a friend said in a conversation I had ingested and a conclusion I had reached about my own work patterns. I had never put those two things in the same sentence. It took four seconds.

gbrain brainstorm "what is the thing I keep avoiding in my thinking about [decision I am sitting on]?"
gbrain lsd "what would have to be true for my plan to [X] to be completely wrong?"

Note on cost: both commands make LLM calls to generate and judge ideas. Each run costs approximately $0.20 to $0.40. Run them with intent, once a month or before any significant decision, not casually.

The output is yours alone. Nobody else’s brainstorm searches your notes.

5. Here Is Where I Land

Three weeks in, the thing I did not expect: the brain made me more deliberate. Knowing that every conversation feeds it made me more careful about what I named out loud, what observations I said to Hermes instead of just thinking them. The input got sharper because the output was permanent.

I have covered this category since Karpathy’s autoresearch post. That was the first system that actually compounded knowledge. A wiki retrieves. GBrain reasons from your history.

The people who start building this now walk into next year with two years of thinking they can actually search. The brain gets more useful the longer it runs. Every conversation adds to it. You do not have to tend it.

Subscribe now

I wrote a complete PM deep dive that hooks GBrain up to your PM OS over on Product Growth. The paid issue includes agent instruction to make any agent get GBrain’s powers.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

That’s today’s deep dive.

But first, a word from our sponsor and the week’s AI news.

LogRocket: The Replay AI That Actually Found My Bug

I ran an eval on my own site this week - landpmjob.com. Same 10 questions, two replay AIs head-to-head: LogRocket Galileo vs PostHog Max.

The gap came from one question: “what’s the highest-impact bug right now?”

Galileo flagged React hydration errors #418 and #423. They hit ~47% of users and broke my Apply Now buttons across header, pricing cards, and FAQ accordions. Max returned zero exceptions. Same site. Same window. (PostHog’s exception capture is opt-in.)

LogRocket auto-captured the most expensive bug on my site without me asking. That’s the eval.

Read the Full Eval

P.S. The full writeup includes all 10 questions, the scoring rubric, and screenshots of both AIs answering each one. If you’ve been wondering whether your replay AI is actually finding bugs or just sounding smart, this is the data.

Try LogRocket

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week's Top News: Google Just Declared the Chromebook Era Over

Fifteen years ago, Google built the Chromebook for a cloud-first world. This week, they announced the Googlebook, and it replaces everything the Chromebook stood for.

But the real story is not the OS. Google's own senior director said it plainly: "We are moving from an operating system to an intelligence system." The Googlebook ships with a feature called the Magic Pointer. You wiggle your cursor over anything on screen and Gemini instantly surfaces what you can do with it. Hover over a date in an email, it offers to create a meeting. Select two images, it composites them together. Point at a flight confirmation, it adds the trip to Calendar. Your cursor has done exactly one thing for 40 years: point and click. Google just turned it into an AI agent.

Bigger picture: this is the most aggressive move anyone has made in the AI PC war. Microsoft spent two years bolting Copilot onto Windows as a sidebar. Apple reserved Apple Intelligence for its own silicon. Google went OS-level from the start, built Gemini into the cursor itself, and got Acer, ASUS, Dell, HP, and Lenovo all lined up for fall 2026 devices. There is even a Glowbar on the lid that lights up when Gemini is active. You will know the difference immediately.

Here’s where I land: the Chromebook worked because it was simple, cheap, and hard to break. The Googlebook is targeting the exact opposite market, premium pricing, premium hardware, and Gemini at every interaction layer. That is a harder sell. But if Google executes and the Magic Pointer actually feels as good in practice as it looks in demos, this is the most interesting laptop category to launch since the iPad blurred the line with tablets.

The Other News That Mattered

• Google launched Gemini-3.5-Flash at Google I/O and it’s one of the most underwhelming releases from a major lab in some time. It’s fast, but it doesn’t move any benchmarks forward.

• SpaceX released its S-1 in which we learned Anthropic is paying SpaceX $1.25B per month for compute. We have a new hyperscaler, and it’s already at $15B ARR.

• Claude for Small Business launched this week with 15 ready-to-run workflows inside QuickBooks, PayPal, HubSpot, Canva, and Docusign. No extra cost beyond your existing subscriptions. Small businesses are 44% of U.S. GDP but have been the last to use AI past the chat window. This is Anthropic’s direct fix for that.

• xAI launched Grok Build, a terminal-based coding agent competing directly with Claude Code. Plan Mode, parallel subagents, 2M token context window. The catch: it’s locked behind SuperGrok Heavy at $300/month. Elon pushed it hard on X, but at that price, most developers are staying put.

• Claude Code shipped agent view in research preview. One unified list of all your active sessions. Small feature, but if you’re running multiple agents at once, it’s the organizational layer that was missing.

Resources

• Thariq’s piece on HTML over Markdown for agent output hit 12.3M views for a reason. HTML renders tables, SVGs, and interactive layouts where Markdown just gives you text. Worth reading if you’re building anything with Claude Code.

• The Claude Code team’s thread on /goal, /loop, and /schedule is the best explanation of how to keep agents running without babysitting them. Set a condition, Claude checks it every cycle, stops when it’s done.

Tools

• Anthropic shipped 12 legal workflow plugins inside Claude Cowork this week, covering commercial counsel, employment law, and litigation, plus MCP connectors to Westlaw and Everlaw. Legal is already the number one job function inside Cowork.

• Codex is now in the ChatGPT mobile app in preview. Start tasks, review outputs, steer execution from your phone while Codex keeps running on your machine.

Funding

• Isomorphic Labs led by Demis Hassabis raised $2.1B in a Series B led by Thrive Capital, with Alphabet, GV, Temasek, and the UK Sovereign AI Fund. The company designs drug molecules from scratch using AI, building on the AlphaFold foundation. They're targeting human clinical trials by end of 2026. If that happens, it's one of the most important things that occurs in science this year.

Share

I Stopped Switching Tabs to Claude. Here’s Why.

The chatbox was never the right interface.

Think about how you actually use AI today.

You’re editing a video. Or writing something. Or trying to figure out a setting inside a tool you’ve been using for years. You have a question. So you leave what you’re doing, open a new tab, go to Claude or ChatGPT, describe what you’re looking at to a model that cannot see it, read the answer, go back, try to find where you were.

You do this twenty, thirty times a day. I counted mine once. It was embarrassing.

The hidden cost is that you’ve been trained to save up questions instead of asking them, because asking them is annoying. Every tab-switch is a small interruption you’ve absorbed so fully you stopped calling it one.

Three years of this, and until this week, I hadn’t seen anyone seriously question the interface itself.

Two Things Happened the Same Week

Farza, possibly the most talented solo builder I know, shipped something called Clicky. It’s a free macOS app that lives in your menu bar, next to your cursor. He built it in three weeks. Here’s the demo.

Here’s what it actually does.

When you hold Ctrl+Option and ask a question out loud, Clicky takes a screenshot of your screen and pairs it with your voice. Both go to Claude at the same time. Claude responds by voice and, when it needs to point at something, it sends a blue animated triangle flying to the exact spot.

You never open a new window. You never describe your screen. You never lose your place.

Unlike Claude where you pick a project, pick a context, figure out which thread applies, here you just speak. Clicky figures out where to route it, what context to pass, and how to respond. You never specify anything.

The onboarding is the best I’ve seen from any AI product. Most companies ship a tour. Farza ships a moment. You feel it working before you’ve finished setting it up.

One thing I want to be honest about upfront: Clicky sees your screen constantly. For personal use that’s fine. For anything work-sensitive, think carefully about that before putting it on a machine with sensitive data. I don’t know how they handle that at scale and I’d want to before using it at a job with strict data policies.

Here’s where I land though: it works. It’s fast. This is the thing I’ve always wanted from AI apps on Mac and didn’t know how to ask for.

Then Google Showed Up With the Same Idea

Two days after Farza’s demo hit three million views, Google DeepMind published research on something they’re calling Magic Pointer.

They called it the first major rethinking of the mouse cursor in more than 50 years. Their framing is almost word for word the same problem Farza solved: AI tools live in separate windows and users have to leave their context to get help. Magic Pointer puts intelligence at the cursor itself. The prompt is a gesture.

Their demos show a user watching a video of Tokyo restaurants, dragging the cursor across a restaurant sign on screen, and Gemini booking a table for that evening. The cursor does the search. The voice does the booking.

This is also shipping to hardware. Google killed ChromeOS and launched Googlebook, an Android-powered laptop where Gemini is baked into the OS and Magic Pointer turns the cursor into an agent. Autumn 2026. This isn’t research anymore.

A solo builder and the world’s best-resourced AI lab, independently, the same week, both answering the same question the same way. That’s not a coincidence. Farza’s answer and Google’s answer are the same: stop making the user go to the AI. Bring the AI to where the user already is.

Setting It Up

1. Go to clicky.so and download the app

2. Install and open it on your Mac

3. Grant screen access and microphone permissions when prompted

4. Hold Ctrl+Option and ask your first question out loud

You're live. No account, no API key, no model selection needed.

One caveat: Clicky is Mac only right now. A Windows version is in the works.

What You Can Actually Do With It

There are two buckets where Clicky earns its keep. Learning tools that normally defeat you, and getting things done without breaking your flow.

Learning tools that used to defeat you

Most complex software has the same problem: the help lives somewhere else. The tutorial is on YouTube. The documentation is in a second tab. The moment you switch away, you lose the context you were in.

Clicky fixes this by being the help, inside the app.

• DaVinci Resolve. One of the most complex creative interfaces on the planet. The old way: YouTube tutorial in one window, app in another, pause, switch, try to find the thing the video just showed, switch back, rewind, repeat until you give up and close both. With Clicky, you stay inside DaVinci, ask out loud how to add a keyframe, and the triangle flies to the exact slider. You try it, fail, ask what went wrong, it explains and points again. The loop of doing, failing, and getting unstuck without leaving the app is how people actually learn things.

• Blender, Figma, Photoshop, After Effects. Same pattern. Any tool dense enough that you’ve ever needed a tutorial running in a second window. Clicky collapses that second window into a voice command.

  

• Any setting you’ve been ignoring. You know the ones. The preference you’ve been meaning to configure for six months. The export setting you never understood. Ask out loud, triangle points at it, you’re done.

Getting things done without breaking your flow

The smaller everyday tasks surprised me more than the learning ones. This is where Clicky starts to feel less like a tool and more like delegation.

• Shopping. The shopping demo felt gimmicky to me at first. But watching Clicky navigate to Amazon, search for resistance bands, and add them to a cart while you’re still looking at your home gym is a different thing to read about than to see. Say “order resistance bands and creatine from Amazon.”

• Reminders. “Remind me about the Sam call at 3pm tomorrow” while you’re looking at whatever triggered the thought. Lands natively in Apple Reminders.

• Messages. Look at your calendar, say “write a WhatsApp message to block time with a client.” Hear the draft. Approve it. Done.

• GIFs. Farza demoed building one from scratch through voice alone. No tool opened manually.

One thing that makes all of this work: the memory holds your last ten exchanges. When you ask “what about the other one,” it knows what you mean. You’re actually having a conversation inside your workflow, not restarting a session every time you have a follow-up.

Why the Eight-Second Interaction Is the Whole Point

Every AI product built in the last three years put AI somewhere you had to go. The sidebar. The tab. The assistant panel. Intelligence was a destination.

What Clicky and Magic Pointer are both showing is that the real interface is your screen. AI doesn’t need its own window. It needs to see what you already see and respond where you already are.

The interactions that keep me reaching for Clicky are the eight-second ones, the quick clarification while I'm already in something. The question I would have saved for later and then forgotten.

Going from two minutes of tab-switching to eight seconds of speaking changes something deeper than time: it changes how often you’re willing to ask for help at all. That’s not what anyone built for. It probably should have been.

I don’t know yet how Clicky handles privacy at scale. I don’t know if Apple builds this natively into macOS and wipes out everyone else who tried it first. I don’t know if Google can ship a cursor-native experience inside their existing products or if it ends up as another panel bolted onto the side.

What I do know: the people using this now will feel the gap every time they switch tabs for the rest of the year.

If you build products or want the deeper take on what the spatial interaction means for AI product design, I go deeper on that later this week in Product Growth.

Subscribe now

1/ A 28-year-old walked into a chiropractor for lower back pain. She left with four dissected arteries, a stroke, and cardiac arrest. It took nine months for her to say “Mom” again. The back pain was four vertebrae below the artery the thrust tore. 1 in 20,000 manipulations triggers this. There is no screening test to identify who’s at risk beforehand. Nobody tells you that before you book the appointment.

2/ Costco has the highest-resolution recession sensor in America. The median Costco household earns above $125K. When that household trades beef for canned tuna, the upper-middle class is admitting their balance sheet changed. A category buyer sees that protein shift in POS data within 48 hours. NBER won’t call the recession official for another 6-21 months. The economists are always last.

3/ Four crocodiles are living in the middle of the Sahara, genetically isolated since the desert was green, somewhere between 5,000 and 15,000 years ago. The water they live in is black with camel dung. The food chain runs camels to dung to algae to fish to crocodiles. Without the herders, the crocodiles starve. Four females. All that’s left of a species that once lived across the entire desert.

4/ When you check in for a flight, one button triggers three systems simultaneously: revenue management locks your seat, TSA runs your name against the No Fly List in real time, and the captain gets your weight for takeoff speed calculations. Check in late on an oversold flight and you move to the front of the involuntary bump queue. I had no idea that button did all of that.

That’s all for today. See you next week,

Aakash

This post is totally free. Share or forward to your friends:

Share

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

They’re more portable, easier to version, and easier to call than prompt libraries.

And with Anthropic’s new SpaceX capacity deal, Claude can handle the token load that skill-heavy workflows demand. Skills have always been powerful. They’re now finally cheap enough to run all day.

Today, I’ll cover how to master Claude Skills. By the end, you’ll have an audit checklist and an eval prompt that hardens any skill you’ve already built. Before that, we’ll have a word from our sponsor and cover the week’s AI news.

Arize: Self-Improving Agents

The teams getting agents to production aren’t relying on manual QA. They’re running evals continuously that let the agent find its own issues and push its own fixes. Arize is the infrastructure layer that makes this possible: instrument your agent, analyze traces, write evals, surface failures, ship improvements. $131M raised. Trusted by the companies actually shipping AI at scale.

Try Arize

P.S. You can get a full year of Arize’s pro plan for free ($1260 value) with my bundle.

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week's Top News: Anthropic Got Elon's Supercomputer

Today, Musk is fighting in court against OpenAI. Simultaneously, he’s renting their biggest competitor the world’s biggest GPU cluster.

Anthropic announced a deal to access all of Colossus 1 in Memphis, 300MW+ of capacity and 220,000 Nvidia GPUs. On the same day, Claude Code’s 5-hour rate limits doubled across every paid plan. Peak-hours throttling is gone for Pro and Max users, and Opus API rate limits went up substantially.

Why did Musk say yes, just days after announcing his acquisition of Cursor?

xAI had already moved training to Colossus 2. Leasing 220,000 GPUs to a direct competitor beats watching them sit dark.

The sentence buried in the xAI announcement that almost nobody is talking about: Anthropic and SpaceX are now exploring orbital compute capacity. Multiple gigawatts. In space. That’s not a roadmap placeholder. It’s a signal about where the infrastructure ceiling actually is.

If you’ve been hitting Claude Code limits, they just doubled.

The Other News That Mattered

• Cognition AI (makers of Devin) released their financials in Colossus magazine, and they’re impressive. $1 million in ARR in September 2024. $445 million run rate today. Usage doubling every eight weeks. Cognition is now raising at $25 billion. That's 56x run rate. Cursor cleared $9.9B at a similar multiple last May, and the multiple held because the curve hadn't bent. Cognition looks similar, with doubling still happening at $445M.

• OpenAI shipped GPT-Realtime-2: the first voice model in the API with GPT-5-class reasoning. It handles interruptions mid-sentence, runs parallel tool calls, and now works across a 128K context window (up from 32K). Two companion models launched alongside it: Realtime-Translate for live speech across 70+ languages, and Realtime-Whisper for streaming transcription. Voice agents finally have the infrastructure to be useful in real production workflows. Here’s the announcement.

• Thinking Machines, the AI startup of former OpenAI CTO Mira Murati, released their first model: a voice model with a 400ms turn-taking latency. That’s the gap between when you stop talking and the model starts. Linguists studying conversation across 10 languages found the universal human average is around 200ms. GPT-Realtime-2 sits at 1630ms. OpenAI and Google have been racing each other on intelligence. Mira built the only model in the human conversation range.

Resources

• Anthropic published Natural Language Autoencoders, a technique that translates Claude’s internal activations into readable text. They used it in pre-release safety testing and caught Claude Mythos Preview cheating on a coding task, then actively trying to cover it up. NLAs showed what the model was thinking before it acted. This is the most concrete interpretability tool Anthropic has shipped publicly. Full thread here.

Tools

• Anthropic shipped dreaming for Claude Managed Agents at Code with Claude this week: a scheduled background process that reviews past sessions, extracts patterns, and consolidates memory between runs. Most agent memory today is a search index. It retrieves facts but never improves from them, which is why agents plateau no matter how much context you give them. Dreaming fixes the compounding problem. My breakdown on why the agents that do this will pull ahead.

• OpenAI made ChatGPT free inside Excel and Google Sheets for all plan users globally, powered by GPT-5.5. Microsoft licenses those same models to power Copilot for M365 at $30/seat/month, roughly $360K/year per 1,000 seats. OpenAI just made the enterprise upsell into the free default, inside Microsoft’s own software.

Share

I am co-hosting the AI Skills Virtual Conf on May 14, free, on Zoom.

30 speakers from Meta, Google, AWS, Scale AI, Bolt, DeepMind, and more. The whole thing is built around practical use cases, not theory. What actually works in 2026, how companies are deciding which AI tools to adopt, and what the real AI stack looks like for founders and small teams.

I am giving a session on how to use Claude specifically for job search: positioning, applications, interview prep, the whole thing.

If any of that sounds useful, register for free here. May 14, 8 AM SF / 11 AM NYC / 4 PM London.

Skills are one of the highest alpha things you can do in AI right now.

I’ve been talking to you about Skills since way before they were cool. My very first edition was on skills way back in October. Since then, I’ve also shipped 6 of my favorite skills for free and a free Claude Setup Skill.

What I haven’t shown you is rigorous testing of what makes a good vs bad skill, and how to improve your own existing skills.

That’s what today’s piece is about:

1. Why Skills Matter

2. How to Deploy Your Skills Everywhere

3. The 7 Laws of Great Skills from 75 Tests

4. How to Improve your Pre-Existing Skills

Subscribe now

1. Why Claude Skills Matter

The 2023 to 2026 meta (“most effective tactic available”) was a prompt library. The 2026 meta is a skill library.

A prompt is text. A skill is behavior.

You have to remember a prompt exists. You have to find the right one. You have to paste it. You re-paste it next session because Claude doesn’t carry state. When you improve a prompt, the old version is still sitting in three other folders. When a teammate uses your prompt, they’re running a version that drifted six edits ago.

A skill is a reusable workflow you install once. Claude loads it automatically when the context matches, without you asking. It knows which files to read first. It knows what the output should look like. It knows what it shouldn’t touch. Standards encoded once apply across every session and every teammate.

The three types of prompt users, and why skills beat all of them

1. The one-liner user. Types “create tickets from this PRD” and hopes for the best. Claude invents the template, the format, and the detail level on the fly. Output changes every session. The skill user types the same thing, and the skill enforces the format, the acceptance criteria structure, the effort estimates, and the dependency map every time.

2. The mega-prompt user. Pastes 800 words at the top of every session. Encodes real standards but burns context tokens before the work starts, drifts with small errors every time it’s re-pasted, and can’t be handed to a teammate as a workflow. The skill user gets the same standards loaded only when relevant, never re-pasted, and shareable as a folder.

3. The structured prompt library user. Has a folder of well-written prompts organized by use case. Better than the first two and still losing to skills on three things: the prompt user has to remember which prompt to invoke, the skill user gets it auto-loaded by description match. The prompt user manually versions across tools, the skill user has one source of truth. The prompt user’s library breaks when switching from Claude to Cursor, the skill user’s library ports.

The skill user wins in every comparison because skills move the work from you to the system. You stop being the router. You stop being the version controller. You stop being the standards enforcer.

When to use a skill vs a prompt

Use a skill when you’d paste the same instructions twice, when the workflow has non-negotiable standards (output format, approval gates, source-of-truth files), when Claude needs to read context before generating, or when you want a teammate to run it identically without a Loom video explaining it.

Leave it as a prompt for one-off questions, exploratory conversations where the prompt is the work itself, simple lookups Claude can answer directly, or tasks that change shape every time.

The test: if you’d paste the same instructions twice, it belongs in a skill.

2. How to Deploy Claude Skills Across Every Surface

The skill architecture is coalescing across Claude, Cursor, Copilot, and Codex.

Make sure you deploy every skill across all the tools you use. Especially all 3 Claude tools. Here’s a quick refresher on how to do that:

Surface 1 - Claude Chat at Claude.ai

Settings, Customize, Skills. Upload the ZIP, toggle on. Type naturally or use /skill-name to force-load.

Surface 2 - Claude Code (terminal)

Personal skills go in ~/.claude/skills/skill-name/ and are active in every project. Project skills go in .claude/skills/skill-name/ inside a specific repo. Commit the folder and everyone who clones the repo gets the same skills automatically. This is the only surface where your skill ships with the codebase, so it’s where engineering standards belong.

Surface 3 - Cowork

Highest leverage for anything document-heavy. The skill reads your actual files, not pasted snippets. Type two words, get output that references your real context. This is the surface where a document too long to paste becomes a two-word command.

Skills also work in Claude for Excel and Claude for PowerPoint.

3. The 7 Laws of a Great Claude Skill

I’ve been using skills every day since October. But over the past week I went further and ran 3 tests on 25 of my top skills across work and personal use cases, trying different skill constructions to see what actually works.

The result is 7 laws of what makes a good skill:

Law 1 - The description is the routing layer

Claude scans every installed skill’s description at the start of a session. The full instructions, the output format, none of it loads until Claude decides the skill is relevant. That decision happens from the description alone.

I sent 10 prompts at a recipe planner skill. The description had 37 characters: “Suggest recipes from what’s in fridge.” Most prompts didn’t trigger it.

A good description is third person, states what the skill does and when to use it in the first sentence, includes trigger phrases a real user would type, and names at least one clear boundary.

The recipe planner skill I tested had a 37-character description. I sent 10 prompts that should have triggered it: “what can I make tonight,” “I don’t want to go grocery shopping,” “help me use up what’s in my fridge.” Most missed.

37 characters is not enough surface for Claude to recognize what the skill does. So the skill stays invisible.

Here’s the same skill with a description Claude can route on:

Law 2 - Every “Do not use for X” needs a “Use /Y instead.”

Most skill writers say when to use a skill. Almost no one says when not to.

When you exclude something without a pointer, Claude still has to pick something. It loads the closest match. The output is wrong format, wrong skill, delivered confidently.

The fix is in the description, not the body. In the body, an exclusion fires after the wrong skill has already loaded. In the description, it fires at routing time.

“Do NOT use for X” is advisory. “Do NOT use for X, use /Y instead” routes correctly.

Law 3 - Write commands, not requests.

Every line in the body should be imperative. “Read the target file. Check for X. Output as Y.” Not “Could you take a look and maybe check for any issues?”

The code reviewer skill I tested opened with: “Hey! When the user asks for a code review, it would be great if you could go through their diff carefully and give them helpful, kind feedback.” Claude matched the energy exactly. The output was friendly, vague, and structured like a message from a thoughtful colleague. No severity ratings. No file and line references:

Switch the body to commands: “Check the diff. Flag every issue with severity (Critical/High/Medium/Low). Reference file path and line number for each. Do not soften.”

Same Claude. Same diff. Different output, because the instructions stopped asking nicely.

Law 4 - Build a read-first table, not just a read-first note.

The biggest single improvement across all 25 skills wasn’t a new rule. It was replacing one vague sentence with one specific table.

The debugger I tested had no read-first step at all. Hand it an error, get a competent-sounding generic analysis. Probable root causes. Suggested fixes. All pulled from training data. None of it touching my actual codebase, error logs, or recent commits.

The fix is three columns: Source, Path, What to extract.

That table tells Claude which directory, which search terms, what to pull out. That’s an instruction. ‘Check the relevant files’ is a label.

Law 5 - Without a template, Claude will invent one.

I ran my daily-plan skill three mornings in a row with the same prompt. Monday: bullet list by time block. Tuesday: narrative paragraph. Wednesday: numbered list with headers.

Same skill. Same prompt. Three different structures.

The instructions were clear. But there was no template and no example. Claude invented a new format every session because I never showed it what done looked like.

The fix is an output template that specifies the exact structure, plus a worked example that shows Claude what a complete output looks like. Rules describe a target. An example shows Claude what hitting it looks like.

Law 6 - One worked example beats five rules.

The commit message skill I tested had 12 rules. Comprehensive. Clear. And inconsistent across three consecutive runs with the same input.

“Concise” means something different every session. There was nothing concrete to pattern-match against.

I added two worked input/output examples and got identical structure across all three runs. Claude is an excellent pattern matcher. Give it a target, not a description of what the target looks like.

If you’re writing more rules than examples, flip the ratio.

Law 7 - Keep skills scoped and short.

Once a skill loads, every line competes with the conversation history for Claude’s attention. The longer the skill, the higher the chance Claude starts ignoring instructions toward the bottom.

I tested a fitness coaching skill that was 724 lines long. Genuinely good content. Training philosophy, injury history, deload protocols. And buried at line 700: the safety rules.

The safety rules never fired.

Move critical rules to the top, before everything else. Move background content to a references folder. Claude won’t read it until it’s needed.

4. How to Improve Your Existing Claude Skills: The Audit + Eval Loop to Steal

I’ve put together all of those laws into an easy Audit Checklist for you:

Description:

• Over 100 characters?

• 3+ phrases a real user would type?

• Written in third person?

• What it does and when to use it in the first 250 characters?

• At least one “Do not use for X, use /Y” boundary?

Body:

• Instructions use imperative verbs?

• Output format specified with a template or worked example?

• Read-first table with Source, Path, and What to extract?

• Critical rules in the first 100 lines?

• Under 500 lines?

Automate It

The checklist is worth knowing. But you don’t have to memorize it. Send this prompt to a chat with your skill:

The checklist gets you to a decent skill. This next prompt gets you to a hardened one. It sets 10 sub-agents loose on your skill, grades the outputs as a hard grader, and rewrites the skill across three rounds based on what broke. I’ve run it on every skill in my library. It surfaces failure modes I’d have never caught by hand.

Paste this into a chat with your skill loaded:

Run it once and you’ll watch Claude find holes you didn’t know your skill had. Run it on your top three skills this week. That’s the highest-leverage hour you’ll spend in Claude all month.

If you’ve just ctrl+a, ctrl+c, ctrl+v’d this into your Claude, it should pick up both prompts.

If you’re a PM, I wrote a version of this specifically for you over at Product Growth. It has a skill file to improve your skills, a skill file to create skills, plus an additional 3 laws for internal knowledge work.

Final Words

Optimize. Your. Skills. Nothing is more powerful than skills in this era.

Run the improvement prompt on your next skill.

You’ll see magic.

That’s all for today. See you next week,

Aakash

This post is totally free. Share or forward to your friends:

Share

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

Everyone is fixating on the growth numbers. The real story is what they are actually using it for.

I’ve been running Hermes for six weeks. Today’s deep dive is your complete guide to what makes it different from every AI tool you already use.

And before that, we’ll have a word from our sponsor and the week’s AI news.

The Chatbot Tax

Last year you added Claude or ChatGPT to Slack. Two people use it daily. Everyone else opens it twice a month. The Monday metrics deck still gets built by hand. That’s the chatbot tax.

Viktor lives inside Slack and Teams alongside the rest of your team, plugs into 3,000+ tools, and ships actual work. A landscaping owner is running 63 Viktor workflows across collections, payroll, and equipment reports. A $10M DTC brand had Viktor build a Klaviyo audit, a 2026 cash flow model, and a checkout debug report in week one. Every action goes through your team for approval first. 11,000+ teams now trust it.

Try Viktor free, $100 in starting credits.

Try Viktor

There’s a billion AI news articles every week. Here’s what actually mattered:

The Week’s Top News: Claude Just Moved Into Your Creative Tools

Claude launched connectors for Blender, Autodesk Fusion, Adobe, Ableton, SketchUp, Splice, and more this week.

Most people I know still think of Claude as a writing and coding tool. That changed this week. If you spend your days in design software, 3D tools, or a DAW, Claude is now inside those tools with you. Not in a separate tab. Not something you copy-paste between. Actually inside your workflow, acting on your real files.

The Autodesk Fusion integration is the one that stopped me. You describe a change to your 3D model in plain language and Claude executes it inside Fusion. That is it. The thing that used to require knowing which menu, which panel, which parameter, now requires knowing what you want. That is a different kind of tool.

The Blender one I keep thinking about for a different reason. Blender has had a full automation layer since 2011. Anyone who has tried to use it knows the problem: you needed someone who could write Python AND knew Blender’s internals well enough to actually do something useful with it. That person is hard to find and expensive to hire. That barrier is now a chat window. Rename objects, fix UV maps, batch changes across an entire scene. No code required.

Here’s where I land: Anthropic announced patron-level support for the Blender Development Fund alongside the connector launch. €240K/year, dedicated to Blender’s Python API, the layer the connector itself depends on. Then today, after community pushback, Blender reversed course and accepted the funds as a one-time donation instead. Anthropic agreed to it.

Read that sequence carefully. The Blender community is wary of AI companies funding the tools they depend on. Anthropic still wrote the check. The integration is shipping either way. The bet on open-source 3D infrastructure is real even when the donation structure changes shape.

The connector is the thing that matters here, not the funding mechanism. Anyone who works in 3D should try it this week.

The Other News That Mattered

• Sundar Pichai announced that Gemini now generates and downloads Word docs, Excel sheets, PDFs, and presentations directly from chat, for free, for all 750M Gemini users. Microsoft charges $360/user/year for Copilot to do the same thing. Only 3.3% of Microsoft’s 450M M365 users have paid for Copilot so far. I wrote about why this breaks the entire Copilot pricing thesis and why the market has not caught up yet.

• OpenAI and Microsoft rewrote their deal and the AGI clause is gone. The provision that would have let OpenAI exit its Microsoft obligations the moment it declared AGI has been removed completely. OpenAI still pays Microsoft 20% of revenue through 2030 regardless of any technical milestone. My read: Microsoft gave up leverage and got certainty on a $135B stake. The quiet winner is Amazon, which can now formally take on the OpenAI workloads Azure exclusivity was blocking.

• Cursor launched its SDK this week. The same agents that power the Cursor app can now run inside your own products, pipelines, and backend systems. Rippling and Notion are already using it to catch bugs and open fixes automatically without anyone prompting anything. Cursor crossed $2B ARR. The SDK is how they stop being a product you open and start being infrastructure you run on.

• xAI launched Grok Imagine Agent Mode (beta) with zero announcement. One prompt, one infinite canvas, and the agent plans, generates, edits, and iterates on its own until it is done. Films, manga sets, product ads. Users just started posting it.

Resources

• Two Anthropic engineers broke down exactly how to prompt Claude for better outputs. If you use Claude every day, this is worth 10 minutes. Most people are still prompting it like a search engine.

• If you use Claude Code and keep running into limits, read this first. Pawel Huryn mapped the four reasons people hit limits that have nothing to do with Claude, with copy-paste fixes for each. The limits are more generous than people assume. The harness is usually the problem.

Tools

• The Claude Blender connector deserves its own mention beyond the main news. The thing that makes it worth paying attention to is not what it does. It is what it replaces. The bottleneck in 3D workflows was never Blender’s capability. It was always finding someone who could speak both Python and Blender fluently at the same time. That skill set is gone as a requirement now. Anyone who works in 3D should try this week.

Share

The AI Agent That Doesn't Forget You

Look at that blue line.

Hermes, from Nous Research. Released February 25. 100K GitHub stars in seven weeks. Six weeks of running it later, here is what makes it different from every AI tool you already use.

In January 2026, a security audit found 512 vulnerabilities in OpenClaw, the agent most developers had been running. One of them let an attacker take full control of your machine through a single malicious link (CVE-2026-25253). Developers left. But the ones who switched to Hermes kept telling me the same thing: they did not stay because of the security. They stayed because it was the first agent that kept working after they closed their laptop.

Here is the thing every AI tool you use today has in common. It only works when you are there. You open ChatGPT, ask a question, get an answer. You close it, and it stops. Nothing happens until you come back and ask something again.

Hermes is built around a completely different idea. It runs in the background, executes tasks on a schedule, and gets better at those tasks every time it runs them. You set something up once and it keeps going. Week after week. Without you.

That is the thing no other AI tool does. And it is the thing I want to show you today.

Here’s what you need to know:

1. What makes this possible

2. Setting it up (20 minutes, then it runs forever)

3. Use case 1: A weekly briefing that arrives before you wake up

4. Use case 2: A health coach that actually knows your patterns

5. Use case 3: Generate images from your phone

6. Three honest limitations

Subscribe now

1. What makes this possible

Two things sit underneath everything Hermes does.

SOUL.md is a short file you write once. It is your standing brief to the agent. Who you are, what matters to you, how you like answers delivered. Mine has things like: I have two kids, my wife is vegetarian, I follow F1 and the NBA, keep answers short unless I ask otherwise.

You write it once. It loads automatically before every session, every scheduled task, every message it sends you. Forever. You never explain yourself again.

The learning loop. Every time Hermes runs a task, it evaluates what worked and saves an improved version of the procedure. The briefing it sends you in week six is sharper than week one. Not because you changed anything. Because it has been running the same task six times and quietly improving each time.

ChatGPT Tasks can fire a prompt on a schedule. That is where the overlap ends. The procedure never improves between runs. The same query goes out Sunday after Sunday with no memory of what worked last week, no skill stored from the last successful execution. And it lives inside ChatGPT. If you want it to reach you in WhatsApp, Telegram, or Slack, you build that yourself. Hermes is the first consumer agent built around the idea that the value compounds the longer it runs.

2. Setting it up

Hermes runs on Mac, Linux, and Windows (WSL2). One command installs everything:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes setup

The installer detects your OS, sets up Python if needed, and walks you through the rest. About five minutes.

Step 1: Pick a model.

Free Option A: NVIDIA NIM. NVIDIA hosts over 100 open-weight AI models in the cloud at build.nvidia.com/models. No GPU. No credit card. Create a free developer account, click any model, and click Get API Key.

hermes model
# Select: Custom endpoint
# Base URL: https://integrate.api.nvidia.com/v1
# API key: nvapi-xxxxxxxxxxxx
# Model: meta/llama-3.1-70b-instruct
# Context length: 65536

Free Option B: Ollama. Runs entirely on your laptop. Nothing leaves your machine. One line most people miss — Hermes needs a 64K context window minimum and Ollama defaults to 4K. Skip this and things break silently:

OLLAMA_CONTEXT_LENGTH=65536 ollama serve
ollama pull llama3.1:8b
hermes model
# Base URL: http://localhost:11434/v1
# API key: ollama
# Model: llama3.1:8b
# Context length: 65536

Already paying for Claude, GPT-4o, or Gemini? Run hermes model and pick your provider. No new cost.

hermes model
# Select: Anthropic → uses your ANTHROPIC_API_KEY
# Select: OpenAI   → uses your OPENAI_API_KEY
# Select: Google   → uses your GOOGLE_API_KEY

Step 2: Connect your phone.

This is the step most people skip. Without it, Hermes only exists when you are sitting at your terminal. With it, it reaches you anywhere.

Hermes works with WhatsApp, Telegram, Slack, and Signal from a single gateway process.

WhatsApp needs no developer account.

During hermes setup, select WhatsApp. A QR code appears in your terminal window.

Open WhatsApp on your phone. Go to Settings → Linked Devices → Link a Device. Point your camera at the QR code. Done. Your agent now lives in your WhatsApp contacts.

# In ~/.hermes/.env — restrict to your number only:
WHATSAPP_ALLOWED_USERS=+1XXXXXXXXXX

I link my personal WhatsApp number directly. Once paired, the agent shows up under Linked Devices exactly like WhatsApp Web.

Telegram is faster if you just want to try it first. Open Telegram, search for @BotFather, type /newbot, follow the prompts, copy the token.

# In ~/.hermes/.env:
TELEGRAM_BOT_TOKEN=your-bot-token
TELEGRAM_ALLOWED_USERS=your-telegram-user-id

Send /start to your bot. Five minutes total.

Step 3: Start the gateway.

hermes gateway

This keeps Hermes connected to your phone. Right now it only runs while your terminal is open. To make it run permanently, surviving reboots:

hermes gateway install

This is the line that makes the Sunday 7am briefing actually land at 7am. Without it, your agent stops the moment you restart your computer.

Step 4: Back up before you touch anything.

hermes backup --quick

Migrating from OpenClaw?

hermes claw migrate

Hermes detects your OpenClaw installation automatically. Everything transfers except API keys, which it skips deliberately for security.

3. Use case 1: A weekly briefing that arrives before you wake up

Every Sunday morning, before my family is up, my WhatsApp gets a message. I did not send it. I did not schedule it manually each week. I set it up once and it has arrived every Sunday since.

This is what it looks like:

I set this up in one message to Hermes:

“Every Sunday at 7am, check the school calendar, the week’s weather, F1 and NBA schedules, and flight prices to Miami. Send me a summary on WhatsApp.”

That was the entire setup. It has been running every Sunday since. I have not touched it.

ChatGPT Tasks can deliver a Sunday briefing. What it cannot do is get sharper at it. Run a ChatGPT Task for eight weeks and week eight is the same as week one. The prompt is fixed. The procedure does not improve. There is no record of which sections you read first or which ones you skipped. Hermes watches all of that and adjusts.

And here is what changes over time. In week one the briefing was basic. By week eight it had learned I care about flight prices specifically in the second week of each month, that I want sports results only for teams I mentioned, and that I like the weather section first because I am getting dressed when I read it. It figured all of that out from watching how I responded each week.

Week one: useful. Week eight: it feels like something that knows your life.

4. Use case 2: A health coach that actually knows your patterns

Everyone has tried a fitness app. You log your workout for four days, miss one, the streak breaks, and you never open the app again. The apps track. They do not think.

I started messaging Hermes my health data the same way I would text a friend. After a run: “30 min run, felt good.” After a bad night: “Slept 5 hours, woke up twice.” After skipping the gym: “Missed today, long day at work.”

After five weeks of this, on a Sunday evening, my WhatsApp buzzed with something I did not ask for:

I did not ask it to find that pattern. I did not ask it to message me proactively. It had been watching 34 log entries across five weeks, connected the skip days to the meeting days, and surfaced it before the day even arrived.

That is the thing no app does. Apps record. Hermes watches.

By week six, it had learned that I sleep better when I stop eating before 8pm, that my workouts are longer on days I have had at least seven hours of sleep, and that I almost always skip Friday exercise if I skipped Wednesday too. It sent me a nudge on Wednesday: “You skipped Monday. Based on your pattern, skipping today makes Friday very unlikely too.”

A health app shows you your data. Hermes tells you what your data means before you need to ask.

This is the use case where compounding is most visible. Month three is genuinely more useful than month one because it knows you better. It has more data, more patterns, more history to work from. A price alert fires once. A health tracker that has been watching you for twelve weeks never stops getting sharper.

5. Use case 3: Generate images/videos from your phone

Nous shipped a ComfyUI skill last week. ComfyUI is the most powerful open-source media generation tool out there, but using it normally means sitting at a computer wiring nodes together every time you want an image.

One install:

hermes install comfyui

Now you message Hermes the image you want from WhatsApp. It builds the workflow, runs it on your laptop, sends the result back. No app to open. No nodes to wire. No data leaving your machine.

The difference from ChatGPT image generation: it runs on your own hardware. Every image is free and private after the local setup. The difference from running ComfyUI directly: you do not have to be at the computer.

6. Three honest limitations

Your machine has to be on. Scheduled tasks need your laptop running when they fire. If it is closed at 7am Sunday, the briefing waits until you open it. macOS has a setting called Wake for Network Access that wakes the machine automatically before a scheduled job. Worth enabling. The other option is running Hermes on a $5 VPS, which is the path I would recommend if you want any of this to actually work without you babysitting it.

The learning loop can learn the wrong thing. It pattern-matches on what you respond to. If you accidentally reply enthusiastically to one bad briefing, it doubles down. There is no “unlearn” command. You either edit the underlying skill file directly or wipe the memory and start over.

Privacy depends on your model choice. The Sunday briefing knows my kids’ school schedule. The health coach has 34 entries of my sleep and workouts. If you point Hermes at GPT-4o or Claude, that data is going to OpenAI or Anthropic on every run. Run Ollama locally if that bothers you. You give up some quality, you get full privacy.

That’s it for today’s deep dive.

I wrote the full version for PMs with three product-specific workflows and a starter kit you can drop in and run today. it’s at news.aakashg.com.

I am co-hosting the AI Skills Virtual Conf on May 14, free, on Zoom.

30 speakers from Meta, Google, AWS, Scale AI, Bolt, DeepMind, and more. The whole thing is built around practical use cases, not theory. What actually works in 2026, how companies are deciding which AI tools to adopt, and what the real AI stack looks like for founders and small teams.

I am giving a session on how to use Claude specifically for job search: positioning, applications, interview prep, the whole thing.

If any of that sounds useful, register for free here. May 14, 8 AM SF / 11 AM NYC / 4 PM London.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

1. OpenAI shipped ChatGPT Images 2.0 - the first image model that actually holds up in a real work workflow. Clean typography, precise layouts, 2K resolution. The memes were everywhere but the real story is what it does with dense documents and UI.

2. Anthropic shipped Claude Design. Not an image generator. Something different. You describe what you want to build, and it hands you back something people can actually click through and navigate in their browser.

Both are worth understanding. The news section covers Images 2.0. Today’s deep dive is your complete guide to Claude Design.

Subscribe now

Arize: Close the Eval Gap

If you are building with Claude Code, Arize just shipped Skills that give your agent native knowledge of eval workflows. Install once - Claude Code already knows how to export traces, run experiments, and optimize prompts.

npx skills add Arize-ai/arize-skills --skill '*' --yes

That is the whole installation. One command and your Claude Code agent has native knowledge of the Arize eval stack.

Try Arize

P.S. You can get a full year of Arize’s pro plan for free ($1260 value) with my bundle.

There’s a billion AI news articles every week. Here’s what actually mattered.

The Week’s Top News: ChatGPT Images 2.0 is the best image model anyone has shipped

OpenAI launched ChatGPT Images 2.0 this week.

Everyone is fixating on the visual quality. But the real story is precision. Prior image models fell apart the moment you asked for anything dense. Small text turned to mush. Logos warped. UI elements became suggestions. Images 2.0 holds all of it at up to 2K resolution. They demoed a full math explainer with clean typography and stepped annotations. Zoom into the rice in their other demo. It holds down to individual grains.

The other thing that changed: it actually knows things now. Updated knowledge cutoff of December 2025. And when you pair it with a thinking model in ChatGPT, it searches the web in real time, generates multiple variations from a single prompt, and double-checks its own outputs. That’s the model doing the work between your idea and the final image.

Four things worth trying this week:

Pair it with thinking mode first. When you select a thinking model in ChatGPT, Images 2.0 searches the web for real-time information, generates multiple distinct images from one prompt, double-checks its own outputs, and can even create functional QR codes. You're not just prompting an image generator. You're briefing something that researches, drafts, and verifies before it hands you the final visual. Try: "Create a visual breakdown of [topic] using the latest information from [URL]."

For time-sensitive work, pair it with thinking mode. It searches the web before generating so you don’t have to manually feed it information. Try: “Make a visual breakdown of [topic] using the latest information from [URL].”

For non-English content, every prior image model mangled non-Latin scripts. This one renders Japanese, Arabic, and Hindi coherently and naturally inside the image. The Japanese manga demo they shipped is the clearest proof. Try: “Create a [social post / infographic] in Japanese for [topic].”

For document-style visuals, it can hold coherent paragraphs of text inside an image without degradation. Useful for anything that needs to look like a handwritten note, a letter, or a detailed written asset.

The Other News That Mattered

• GPT-5.5 launched in ChatGPT and Codex, positioned as a new class of intelligence for agentic work. The model story is interesting but the pricing tells you everything. GPT-5 launched at $0.63 per million input tokens. GPT-5.5 is $5.00. That’s an 8x increase in eight months. Every developer paying 2x for inference is directly funding the products now competing with their own. Think about that for a second.

• Workspace agents launched in ChatGPT, shared teammates that run long workflows across tools and teams, even when you’re offline. This puts the same pattern Notion built in front of 900 million weekly users and 7 million paid workplace seats. My take: every mid-tier SaaS AI SKU just got 18 months shorter.

• SpaceX secured the right to buy Cursor for $60B later this year, or pay $10B to walk away. The strategic logic is Colossus: a million H100-equivalent GPUs that need a software reason to exist. Elon bought a one-year call option on the best coding AI in the market. That’s the whole deal.

• Claude connected to 15 new apps this week: Tripadvisor, Booking.com, Instacart, Spotify, TurboTax, Uber, Thumbtack, and more. Every single company on that list built their business by aggregating supply. Now AI is aggregating them. And they’re lining up to let it happen. Thumbtack handed Claude access to 300,000+ service businesses. Intuit connected TurboTax. These aren’t API partnerships. They’re customer relationship transfers

Resources

• If you’re tracking whether LLMs are the endgame, read this. LeCun left Meta, started AMI Labs, and just shipped LeWorldModel - 15M parameters, single GPU, 48x faster planning than foundation-model alternatives. One regularizer replaced years of engineering hacks. It learned physics without being told physics existed.

• If you’re deploying agents at work, send this to your security team now. Google DeepMind mapped six ways to hijack AI agents - not by touching the model, but by corrupting what it reads. Exploit rates hit 86% in tests. The attack surface isn’t the model. It’s the internet.

Tools

• Cowork now builds live dashboards connected to your apps and files. Tableau is $75–$115/user/month. Looker starts at $30K/year. Claude Pro is $20/month. The BI seat is starting to compress into a chat message.

• Memory on Claude Managed Agents is in public beta. Agents now learn across sessions - tracking tasks, preferences, and context over time. Stored as files, fully exportable, API-managed. This is what makes agents actually useful instead of starting from zero every time.

Share

Claude Design: You Can Now Go From Idea to Something Real Without a Designer

The complete guide to building prototypes, presentations, and launch pages from a prompt.

I have spent years watching people get stuck at the same point. They have an idea. They can describe it perfectly. But turning that description into something they can actually show someone costs money they do not want to spend, time they do not have, or a dependency on someone else’s calendar.

Anthropic just shipped a tool that solves exactly that. You describe what you want. It builds a clickable, shareable, fully interactive version of it in minutes. No design skills. No software to learn. No waiting on anyone.

I have been testing Claude Design this week. Here is what it is and what you can build with it tonight.

Here’s what you need to know:

Today’s Deep Dive

1. What Claude Design actually is

2. How to get started - three steps before you build anything

3. Four things you can build tonight

4. How to fix the first draft

5. Where I land

Subscribe now

1. What Claude Design Actually Is

Claude Design is Anthropic’s new tool for building visual things without design skills. You describe what you want. It builds a fully interactive, clickable version in the browser. You refine it by talking to it. When you are done, you share a link.

No design software. No waiting for someone else to have availability.

It is included in your existing Claude Pro, Max, Team, or Enterprise plan. Nothing new to buy.

Here is what the output actually looks like. You describe an idea - a landing page for your side project, a presentation for a meeting, an interactive mockup of something you want to build. Claude Design generates it as a live page in the browser. Every button works. Every screen navigates. You share a link and the person on the other end clicks through it like a real product.

What surprised me was what happens after the first draft. You do not restart when something is wrong. You talk to it. “Make the hero section darker.” “The second screen needs a progress bar.” “Change the font.” It adjusts. The canvas is persistent. It remembers everything you have built in the session.

The one thing that changes output quality more than anything else:

Before Claude generates, it asks you four clarifying questions about visual direction, tone, how adventurous the design should feel, and brand personality.

Most people click through these quickly without thinking. That is the mistake. Pick something bolder than feels safe. You can always pull it back. I did this wrong in my first two sessions. The outputs were technically correct and completely forgettable. When I changed my answers, the first draft was something I actually wanted to show people.

2. How to Get Started - Three Steps Before You Build Anything

Step 1: Go to claude.ai/design

You need a Claude account. Pro plan ($20/month) gives you access. If you already use Claude, you already have it - nothing new to install or sign up for.

When you land, you will see four tabs across the top: Prototype, Slide deck, From template, and Other.

• Prototype is for anything interactive - a page, an app idea, a flow someone can click through

• Slide deck is for presentations you want to export as PowerPoint or Canva

• From template has one built-in option right now: an animation template for motion-based pages

• Other is for anything that does not fit the first two

If you are not sure which one to pick, start with Prototype. It covers almost everything.

Step 2: Pick High Fidelity, not Wireframe

Inside Prototype, you will see two options: Wireframe and High Fidelity. Always pick High Fidelity. Wireframe gives you a rough sketch with placeholder boxes. High Fidelity gives you something that looks designed and feels real. The whole point is to have something you can actually show people.

Step 3: Answer the clarifying questions properly

Before Claude builds anything, it asks you four questions about visual direction and tone. Most people click through these quickly. That is the single biggest mistake you can make.

The question that matters most is how adventurous you want the design. Most people pick “By-the-book” and then wonder why the output looks like every other generic page. Pick something bolder. If it is too much, you can tone it down with one click in the next step. I did this wrong in my first two sessions - the outputs were technically correct and completely forgettable. When I changed my answers, the first draft was something I actually wanted to show people

3. Four Things You Can Build Tonight

A Page for Your Side Project, Newsletter, or Small Business

You have something worth sharing and no clean way to share it. A half-finished Notion page. A Google Doc you keep meaning to turn into something real.

I built a full product page in 10 minutes. Headline, feature grid, social proof section, a call to action. Something that would have taken a designer half a day to commission. I described it in a paragraph and Claude Design built it.

→ Tab: Prototype → Mode: High Fidelity

Your prompt:

Build a landing page for [what you do or make].

Audience: [who it is for]

Three things the page needs to communicate:
1. [what it is in one sentence]
2. [why someone should care]
3. [what you want them to do next]

Sections: hero, 3 key points, a simple call to action

Visual direction: [clean and simple / bold / editorial / warm]
Tone: [professional / friendly / energetic]

→ Export: Share → Export as standalone HTML. You get a live URL to share or embed anywhere.

A Presentation That Looks Like You Spent All Week on It

Most people give presentations with slides that signal they ran out of time. Claude Design builds the deck. You write the content.

I built a 10-slide deck in 8 minutes. Exported it as PowerPoint. Speaker notes already written because I turned the speaker notes toggle on before generating. Claude put the detail in the notes and kept the slides clean.

→ Tab: Slide deck

One setting to check before generating: There is a “Use speaker notes” toggle. If you are presenting live, turn it on. Claude puts less text on each slide and writes the spoken detail into speaker notes automatically. If the deck is something people read on their own, leave it off.

Your prompt:

Build a [X]-slide presentation for [audience - a team meeting, a client, a job interview].

Topic: [what this is about]

Slide by slide:
- Slide 1: [title + one-line summary]
- Slide 2: [the problem or context]
- Slides 3-7: [your main points, one idea per slide]
- Slide 8: [results, evidence, or a demo]
- Slide 9: [what you are asking for]
- Slide 10: [close + contact]

Tone: Confident. No filler slides. Every slide earns its place.

→ Export: Share → Export as PPTX for PowerPoint or Keynote, or Send to Canva if you want to keep editing it there.

A Launch Page for Something You Are About to Ship

Side projects, newsletters, events, products. Everything that deserves a real page but cannot justify the cost of one.

I built a launch page with a proper hero section, a before/after comparison, and a call to action in 15 minutes. Then I went further and added an animated video background. That version took 45 minutes total, including generating the video.

For the animated version: generate a short looping video in Kling on the free tier. Upload it to Mux to get a live streaming URL. Then tell Claude Design to use it as the hero background. You can add motion effects by pasting pre-built animation prompts from motionsites.ai - free prompts that handle the CSS and animation work. I tried writing the animation code myself first. Twenty minutes and mediocre results. The pre-built prompt worked on the first paste.

Your prompt:

Build a single-page launch page for [product or project name].

What it does: [one sentence]
Who it is for: [specific person]
What makes it different: [one thing]

Sections: hero, 3 features or benefits, one proof point, CTA

Visual direction: [clean / dark / bold / warm]

Show Someone Your Idea Before You Spend Any Money on It

Before you commission anything. Before you pay a developer. Before you commit to a direction. You can share a fully clickable version of your idea and get real feedback on whether it works.

The moment someone sees “John Smith” or a placeholder price ending in “.00” they stop reviewing your idea and start reviewing why it does not feel real. Before I share any prototype, I go through every field with the Edit tool and replace placeholder data with something plausible. Real-sounding names, realistic numbers, actual copy for the domain. It takes five minutes and it completely changes the quality of feedback you get back.

Your prompt:

Build an interactive mockup of [your idea].

Who will look at it: [a friend / potential customer / collaborator / investor]
What I want to know from them: [one specific question]

Screens:
- [Screen 1 name]: [what happens here]
- [Screen 2 name]: [what happens here]

Every button should navigate. No placeholder text.
Make the sample data feel real for [your industry or context].

→ Share: Share → Copy link. The person opens it in a browser. No account. No download. They just click through it.

4. How to Fix the First Draft

The first draft will not be perfect. That is not a flaw. Claude Design is built for iteration. Here is the order that changes everything.

Tweaks first. Always.

The first draft almost never matches your intended colors or visual style exactly. Tweaks fixes this globally in one move. One adjustment changes every screen at once - the accent color, the illustration style, the background tone. It is completely free. It takes 90 seconds.

I ran Tweaks on every session I tested. In my testing it closed about 60% of the gap between the default output and what I actually needed. Run through every option before touching anything else.

Edit second.

After Tweaks, small things will still be wrong. A headline that is off. A button label that does not match. Sample data that looks fake.

Click the element directly and change it. Three seconds, no cost. The mistake I kept making early on was writing a comment asking Claude to fix something I could just click and type. Comments cost tokens. Clicking and typing does not.

If you can click on it, always Edit it directly. Never write a comment for something you can fix by clicking.

Comments last.

Comments are for when a whole section needs to be rebuilt. A screen that is in the wrong order. A layout that does not work. Something that requires Claude to regenerate rather than you clicking to fix.

Write a specific note, select the section it applies to, and batch everything before sending. Three changes in one send, not three separate sends. Each send is a full AI call. “Make the second screen show the results before asking for payment” works. “This feels wrong” does not give Claude enough to act on.

5. My take on using Claude Design daily

The thing I did not expect was how much the sharing changes things.

When you have something clickable to share instead of a document to read or a Loom to watch, the conversation around it is completely different. People engage with what is in front of them. They click around. They notice specific things. The feedback gets useful because the thing they are responding to is specific.

That shift from “I have an idea” to “here is a link, click through it” used to require hiring someone or learning a tool that takes months to get good at. Claude Design makes it a 15-minute conversation.

The quality ceiling is real. What Claude Design produces is a strong first draft, not a finished product. For something going into production, a designer will take it further. But for getting to the point where you have something real to react to - something to show a collaborator, get feedback from a potential user, or use to make a decision - the first draft is more than good enough.

Two honest things before you start:

1/ Token costs will surprise you. A heavy session with multiple rounds of Comments can burn through your weekly Pro plan quota fast. If you plan to use this regularly, Max tier is worth it. A complete prototype or 10-slide deck costs roughly $2–7. That math works if you are replacing a contractor or a back-and-forth cycle. Just know it before you start, not when you hit the limit mid-session.

2/ This is still the beginning. Everything about Claude Design right now is going to look primitive in 18 months. I genuinely cannot wait to see where this goes. The people who are using it now and developing judgment about what works are going to have a real advantage when the tool gets dramatically better. Get the reps in now.

PS. The Product Growth edition covers everything from the PM side - including a starter GitHub repo. If that’s useful for you, it’s at news.aakashg.com.

Beyond AI, here’s what I found interesting this week:

1/ Apple named John Ternus as next CEO. He joined straight out of college in 2001 and has never worked anywhere else. The board passed on the software guy, the services guy, and the actual chip designer. Tim’s bet was supply chain. John’s bet is the stack.

2/ Rockstar spent $256M to build GTA 5. It’s now done $10B in revenue, still pulls $500M+ a year in year 13, and players are still finding things. The math only works when the simulation has more depth than any player will ever exhaust.

3/ Movie tie-in games are dead because HD broke the math. In the PS2 era, THQ shipped Pixar tie-ins for under $10M and booked real profit. HD multiplied asset budgets while deadlines stayed tied to movie release dates. Nobody runs that trade anymore.

4/ The GTA V train isn’t actually moving. Rockstar scripted it to teleport its coordinates along a fixed path every frame. It doesn’t break your tanker. It doesn’t know your tanker exists.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

Then, the deep dive this week is how 4.7 + two other Claude releases - Claude Routines and Claude Managed Agents - creates magic.

You’re not going to want to miss this one.

Arize: Close the Eval Gap

Everyone’s building AI products. Almost nobody is evaluating them.

Arize is the platform that closes that gap. Their open source tool Phoenix has 2M+ monthly downloads. Their enterprise platform Arize AX is used by Uber, Booking.com, Duolingo, and hundreds more.

One command in Claude Code and your agent is fully traced to Arize. You can see every decision point and every failure, and this data is the foundation for improving your agent.

Try Arize

P.S. You can get a full year of Arize’s pro plan for free ($1260 value) with my bundle.

There’s a billion AI news articles every week. Here’s what actually mattered.

The Week’s Top News: Anthropic Shipped a Weaker Model on Purpose

Anthropic released Claude Opus 4.7 on Thursday. It’s the best publicly available model for agentic coding. They also put the model they decided was too dangerous to release to the public in the same comparison table and showed you the gap on a chart.

Mythos Preview is gated to defensive security teams and roughly 40 companies. If you’re reading this as a builder, that rightmost column is not your reality. It’s the ceiling, not the floor. Everything else in this piece is available to you today.

Quick decoder for the table below: SWE-bench measures whether an AI can fix real GitHub issues autonomously. Terminal-Bench tests command-line coding tasks. Humanity’s Last Exam is a set of questions designed to be too hard for any AI. The percentages are success rates - higher is better.

This is the most unusual launch chart in frontier AI history.

Look at the rightmost column. Mythos Preview beats Opus 4.7 by 13 points on SWE-bench Pro, 10 on Humanity’s Last Exam, 13 on Terminal-Bench. Per Anthropic’s Mythos Preview system card, the model developed working exploits for patched Firefox vulnerabilities 181 times in the CyberGym eval, compared to 2 for Opus 4.6. Anthropic decided those capabilities were too dangerous to ship, gated Mythos to defensive security teams only, then put it in the comparison table of the model they will sell you.

That’s not how any other lab operates. OpenAI makes the released model look like the frontier. Google does the same. You’re never supposed to know what they’re sitting on. Anthropic just told you exactly what they held back, why they held it back, and showed you the gap on a chart.

The gap between those with the most powerful models and those without continues to expand.

The Other News That Mattered

• Perplexity rebuilt Mint by adding full Plaid bank integration inside Perplexity Computer. Mint died in 2024 because nobody would pay for budgeting as a standalone product. Perplexity already has the $20/month subscription. Finance is just a retention feature now, not a business model. That’s the structural reason AI companies keep eating vertical SaaS.

• Shopify gave every AI coding agent direct write access to 5.6 million store backends through the new AI Toolkit. Products, orders, inventory, SEO, images. One prompt now does what used to cost $200 to $500 a month in apps plus a $2,000 SEO audit. Shopify isn’t building the agent. They’re building the protocol that makes every agent a Shopify agent. That’s the platform play.

• Claude Code shipped /ultrareview this week alongside Opus 4.7. When you run it, Claude spins up multiple AI reviewers in the cloud that check your code in parallel. Every bug one reviewer flags gets independently confirmed by another before it shows up in your results. No more reviewing the reviewer.

• Google launched Skills in Chrome on desktop this week. You save a prompt once, hit forward slash, and it runs on any page you’re viewing. Macro automation for your browser, no code required. It’s the kind of feature that sounds small until you realize you’ve been retyping the same prompts into Gemini every single day.

• Windsurf 2.0 launched with Devin baked directly into the IDE. You plan locally, hand tasks to Devin with one click, close your laptop, and Devin opens PRs while you sleep. OpenAI offered $3B for Windsurf. Google paid $2.4B for the CEO and top researchers. Cognition picked up the rest: the IDE, 350 enterprise customers, and $82M ARR, for a reported ~$250M - roughly 3x ARR for a developer tool with enterprise traction (Cognition didn’t disclose terms). That deal is looking better every week.

Resources

• I think you need to start plugin maxxing. Zedchi went through 101 official Claude Code plugins and picked the 36 actually worth installing across 8 categories.

• Dwarkesh released one of the best AI podcasts yet in his combative interview with Jensen Huang. The highlight for me was Dwarkesh pointing out the holes in Jensen’s publicly stated statements. Jensen spends the whole podcast trying to sell his chips. Impressive performance for the 25 year-old Dwarkesh.

Tools

• Google Skills in Chrome lets you save any prompt and run it on any page with one click. Macro automation for your browser, no code required. Rolling out to all desktop users now.

• VoxCPM2 just made studio-quality voice cloning free. 30 languages, 48kHz output, runs on a gaming laptop. ElevenLabs charges $0.16 per minute. VoxCPM2 charges your electricity bill. I wrote about what this means for ElevenLabs’ $500M raise.

Subscribe now

Claude Routines and Claude Managed Agents: The Builder’s Guide

Claude Opus-4.7 made all the headlines. But the real magic is in combining Opus-4.7 with two features Anthropic released earlier in the week: Routines and Managed Agents.

Routines automate your own work. Managed Agents automate your product’s work.

That’s today’s deep dive:

1. Routines vs Agents

2. What this means even if you never write code

3. Claude Routines: what they actually are and builds for every role

4. Claude Managed Agents: the production layer and 3 real deployments

5. What’s live, what’s preview, and what breaks

Subscribe now

1. Routines vs Agents

Before this, building a production AI agent meant two jobs.

Job one: design what the agent does. Job two: build everything that makes it run - the servers, the security, the session management, the tool wiring, the monitoring. If those words mean nothing to you, skip to Section 2 and come back. If they mean everything to you, keep reading: Anthropic just eliminated that entire job.

That second job had nothing to do with your product. It took most teams three to six months. And you rebuilt it from scratch every time a new model shipped, because your harness encoded assumptions about what the old model couldn’t do.

Claude Routines are scheduled Claude Code cloud sessions. Prompt plus connectors plus trigger. No server. No agent loop. No infrastructure. Fires on a schedule or a webhook and runs until it’s done. Available on any paid Claude plan today.

Claude Managed Agents are a hosted production layer for shipping agents to users. You define the agent once - model, system prompt, tools, permissions. Anthropic runs the servers, sandboxing, session persistence, credential vault, tool orchestration, and observability.

Routines are for your own recurring work. Managed Agents are for things you ship to other people.

2. What this means even if you never write code

If you're not a developer, here's why you should care.

Routines connect to Gmail, Slack, Google Calendar, Notion, and the web. That means any recurring task you do with those tools can now run on autopilot: pulling competitor news every morning, prepping a summary before your Monday standup, flagging overdue items in your project tracker, or compiling weekly metrics from multiple sources into a single Slack message. You write what you want in plain English, pick a schedule, and it runs while you sleep.

Managed Agents matter because they change how fast your engineering team can ship AI features. If you're a PM writing a brief for an AI-powered feature, the timeline just collapsed from "3-6 months of infra work before we start" to "days."

The builds below show exactly what that looks like. Even if you skip the code prompts, read the use cases - they'll change how you scope your next roadmap.

3. Claude Routines

What a Routine actually is

A Routine is a saved Claude Code configuration: prompt, connectors, and a trigger. It runs on Anthropic’s cloud infrastructure as a full Claude Code session. Your laptop can be off.

Read this before you write any prompt in this section.

Every Routine run starts from zero. No memory of yesterday. If your Routine needs continuity (what changed since the last run, what issues are still open, what you already flagged), write today’s output to Notion at the end and read it back at the start of the next run. Every prompt below assumes you’re doing this.

Second thing that burns people: all connected integrations are available by default and Claude can use them without asking. Remove every connector the Routine doesn’t need before you schedule it. Less access, less surface area.

Three trigger types

1. Schedule — minimum one hour interval. Daily, weekly, or hourly. The standard case.

2. Webhook / API — HTTP POST to a per-Routine endpoint with a bearer token. Research preview. Good for event-driven flows.

3. GitHub — fires on PR open, push, release. Each matching event starts its own session. Good for code-adjacent automation.

The difference from a cron job running a script: the session reasons. It doesn’t execute a fixed procedure. It reads output, makes judgment calls, handles edge cases mid-run. A bash script can’t decide whether a dependency vulnerability is worth waking someone up for. A Routine can.

Routine 1: Competitor intelligence brief

If you manage a product, run marketing, or lead a team, start here.

Every morning before you open Slack, this Routine has already scanned your competitors’ websites, blogs, and social accounts for anything new. Product launches. Pricing changes. New positioning. Job postings that signal where they’re investing. It writes a short brief and drops it in a Slack channel or Notion page.

Connectors needed: Slack (or Notion). Web browsing is built in.

Go to claude.ai/code/routines. New Routine. Set trigger to Schedule, daily at 7 AM. Paste this:

You are a competitive intelligence analyst.

Every morning:

1. Visit these competitor websites and blogs: [LIST URLS]

2. Search the web for news about these companies from the last 24 hours: [LIST COMPANY NAMES]

3. For each competitor, note:
   - Any new product announcements or feature launches
   - Pricing or packaging changes
   - New job postings that signal strategic shifts (e.g., hiring 5 AI engineers = investing in AI)
   - Blog posts, case studies, or press mentions

4. Read today’s context from: [NOTION PAGE URL or “yesterday’s Slack message in #competitive-intel”]

5. Compare against yesterday’s findings. Only report what’s NEW.

6. Post to Slack channel [CHANNEL ID]:
   Competitive Brief — [Date]
   [Company Name]: [What changed and why it matters in 1-2 sentences]
   If nothing changed: “No new competitor activity in the last 24 hours.”

7. Write today’s findings to [NOTION PAGE URL] so tomorrow’s run has context.

Keep the brief under 200 words. No filler. Only report things a product or marketing lead would act on.

Do not write anything back to me. Execute these steps immediately.

One known limit: web browsing finds publicly available content only. It won’t catch changes behind a login wall or gated announcements.

Hit New Routine:

Paste the Prompt with your details:

And it’s on:

Do not write anything back to me. Execute these steps immediately.

Routine 2: Weekly team digest

If you run a team and spend Monday mornings piecing together what happened last week, this replaces that.

Every Monday at 7:30 AM, this Routine reads your team’s Slack channels, Calendar events from the past week, and a Notion project tracker. It writes a digest: what shipped, what’s blocked, what meetings are coming up, and what needs your attention.

Connectors needed: Slack + Google Calendar + Notion.

You are a team operations assistant.

Every Monday at 7:30 AM:

1. Read messages from the last 7 days in these Slack channels: [LIST CHANNEL IDS]
2. Pull all calendar events from the last 7 days for: [LIST TEAM MEMBER EMAILS]
3. Read the current state of this Notion project board: [NOTION PAGE URL]
4. Write a digest with these sections:
   SHIPPED LAST WEEK: items that moved to Done on the project board
   BLOCKED: items flagged as blocked, with who owns them
   KEY DECISIONS: any decisions made in Slack threads (look for messages with reactions or explicit “decision:” language)
   THIS WEEK’S CALENDAR: important meetings coming up in the next 5 days
   ATTENTION NEEDED: anything that looks stalled, overdue, or unassigned

5. Post the digest to Slack channel [CHANNEL ID].
6. Keep it under 300 words. Use short sentences.

Do not write anything back to me. Execute these steps immediately. 

It pings you in Slack:

Routine 3: PR review on every push

If you ship code daily, this one's for you.

You push code. Nobody reviews it until standup. By then the context is gone.

This Routine fires on every new PR via GitHub trigger. Reads the diff. Checks for obvious issues - missing error handling, hardcoded values, undocumented functions, test coverage gaps. Posts a structured review comment directly on the PR before any human has looked at it.

Not a replacement for code review. A first pass that catches the easy stuff before a senior engineer has to.

Connectors needed: GitHub.

A note on model choice: Use Sonnet 4.6 for this, not Opus. Opus is slower and roughly 5x the cost per token. For a monitoring task that runs on every PR, Sonnet handles the reasoning fine and keeps your bill reasonable.

Go to claude.ai/code/routines. New Routine. Set trigger to GitHub, connect your repo. Paste this:

You are a code reviewer for [REPO NAME].

When triggered by a new pull request:

1. Read the full diff.
2. For every function signature, return type, or exception path that changed in the diff, trace the downstream callers in this repo. Identify any caller where the contract changed but the caller was not updated in the same PR.
3. For every new external call (API, DB, filesystem), check whether the surrounding code handles the failure mode it can actually produce (timeout, auth failure, malformed response). Not "is there a try/except" — does the handler match the real failure.
4. For every changed branch of logic, check whether an existing test covers the new behavior. If not, name the test file that should have been updated.
5. Also flag: hardcoded secrets, N+1 queries, and unbounded loops.
6. Post a review comment on the PR:


- No issues: "Automated review: no contract breaks, failure modes handled, tests cover changed paths. Ready for human review."
- Issues: list them by severity (contract breaks first, then failure handling, then test gaps, then style).

Keep the tone factual. No praise. No padding.
Do not write anything back to me. Execute these steps immediately.

Click Run Now on a test PR first. Confirm the comment appears. Then enable the GitHub trigger.

Scale limit to know: Step 2 ("trace the downstream callers") works well on repos under ~5,000 files. On large monorepos, Claude may time out or miss callers outside the context window. For repos that size, scope the review to the changed files and their direct imports only.

Routine 4: Dependency audit, weekly

If you maintain a production codebase, this catches the things you forget to check.

Your package.json has 47 direct dependencies. You have no idea which ones are outdated, which have known CVEs, or which haven’t been touched in three years.

Every Monday at 7 AM, this Routine reads your requirements file, checks each dependency against current stable versions, flags anything with a known CVE or more than two major versions behind, and posts a ranked report to Slack. Severity one is a security issue. Severity two is significantly outdated. Severity three is worth noting.

Connectors needed: GitHub + Slack.

You are a dependency security analyst for [REPO NAME].

Every Monday at 7 AM:

1. Read the requirements.txt or package.json from the main branch
   of [REPO URL].

2. For each dependency, check:
   - Current version in the file vs latest stable release
   - Any known CVEs against the installed version
   - Last release date of the package

3. Classify each finding:
   SEVERITY 1: Known CVE against installed version
   SEVERITY 2: Two or more major versions behind latest stable
   SEVERITY 3: No release in 18+ months (maintenance risk)

4. Post to Slack channel [CHANNEL ID]:

   Dependency Audit — [Date]

   SEVERITY 1 (fix now):
   - [package] [installed version] — CVE: [ID] — [one line description]

   SEVERITY 2 (update this sprint):
   - [package] [installed] → [latest]

   SEVERITY 3 (note for backlog):
   - [package] — last release [date]

   If nothing flagged: "Dependency audit complete. No issues."

Do not write anything back to me. Execute these steps immediately.

4. Claude Managed Agents

What the architecture actually is

Managed Agents has three components. You need to understand all three before you write a system prompt or brief an engineer.

1. The Agent is the brain. Model, system prompt, tools, MCP servers, permission policies. You define this once. It gets an ID. Every session references that ID.

2. The Environment is the hands. A cloud container with Python, Node.js, and Go pre-installed. Network access rules you configure. Mounted files. Anthropic provisions it, scales it, and tears it down. You define what it can reach.

3. The Session is the log. Append-only event stream. Every tool call, every decision, every output, timestamped and traceable. Persists through network disconnections. Survives for hours.

The brain/hands separation is what makes this infrastructure-safe. Claude generates code inside the session. That code runs in the sandbox. The sandbox never handles raw credentials - API keys are stored in a vault outside it and proxied in via MCP. A compromised prompt can’t exfiltrate your tokens.

Built-in tools available to every agent out of the box: bash, file read/write, web search, web browse, code execution, and MCP server connections.

Pricing: $0.08 per session-hour (idle time is free), plus standard token costs ($3/$15 per million tokens for Sonnet 4.6 input/output). A practical example: a 10-minute Sonnet session that processes 50K input tokens and generates 5K output tokens costs roughly $0.16 in compute plus $0.01 in session time. Most runs cost pennies.

Managed Agents Build 1: Bug-to-PR agent

If your team spends hours going from “bug identified” to “fix in review,” this collapses that to minutes.

This is what Sentry actually shipped.

Their debugging agent Seer identifies the root cause of a flagged bug. Previously that’s where it stopped: a diagnosis with no fix. They paired it with a Claude Managed Agent that reads the diagnosis, writes the patch, and opens a PR. Flagged bug to reviewable fix, zero human steps.

The numbers from Sentry’s writeup: over 1 million root cause analyses processed per year, near-immediate reviews on more than 600k PRs per month, and the initial integration shipped in weeks by a single engineer.

Two things worth noting about how they got there. First, they didn’t build the sandboxing or session layer. Anthropic runs both. Second, they kept the scope tight: the agent only handles bugs Seer has already diagnosed. It doesn’t diagnose. It doesn’t refactor. One job, one handoff, one PR.

The session runs for however long it takes. Could be four minutes. Could be forty. Anthropic handles the runtime either way.

Go to platform.claude.com, Managed Agents, Quickstart. Paste this into the “Describe your agent” box:

You are an autonomous bug fix agent for [REPO NAME].

When triggered with a bug report and root cause analysis:

1. Read the full context: error message, stack trace, root cause
   summary, and affected file paths.

2. Check out the affected files from the main branch.

3. Write a minimal fix that addresses the root cause without
   changing unrelated logic. Do not refactor. Do not optimize.
   Fix the specific bug.

4. Write or update the test that would have caught this bug.
   The test must fail on the unfixed code and pass on your fix.

5. Open a PR against main:
   - Title: "Fix: [error type] in [file]"
   - Body: root cause summary, what changed, how to verify

6. Post to Slack channel [CHANNEL ID]:
   "Bug fix PR opened: [PR URL]. Affected: [file]. Review
   required before merge."

Never merge without human approval. If the fix requires
changes to more than 3 files, post to Slack and stop:
"Fix scope too broad for autonomous PR. Manual review needed."

After launch: the scope limit (”more than 3 files, stop”) is the most important line. Remove it and you’ll get PRs that touch half the codebase. Keep it for the first month.

Managed Agents Build 2: Codebase security scanner

If your security audits are quarterly and expensive, this runs weekly for pennies.

Most security audits are quarterly, expensive, and done by people who don’t know your codebase. This agent runs weekly, knows your whole repo, and costs $0.08/hour.

Every Sunday night, it reads every file in your repo, maps the attack surface, cross-references against known vulnerability patterns, and outputs a structured report ranked by severity and exploitability.

You are a security analyst for [REPO NAME].

When triggered:

1. Read every file in the repository. Focus on:
   - Authentication and session handling
   - Input validation and sanitization
   - Database queries (SQL injection surface)
   - File operations and path traversal risks
   - Dependency imports with known CVE history
   - Hardcoded secrets, tokens, or credentials
   - API endpoints with missing authorization checks

2. For each finding:
   - Location (file and line)
   - Vulnerability class (OWASP category if applicable)
   - Severity: Critical / High / Medium / Low
   - Exploitability: how a real attacker reaches this
   - Recommended fix in one paragraph

3. Create a Notion page titled "Security Scan — [Date]" with:
   - Executive summary (3 sentences)
   - Findings table sorted by severity
   - Full details per finding

4. Post to Slack channel [CHANNEL ID]:
   "Security scan complete. [X] critical, [Y] high, [Z] medium.
   Full report: [Notion URL]"

Flag false positives explicitly rather than omitting them.
A false positive with explanation is better than a missed finding.

Scale limit: “Read every file” works on repos up to a few thousand files. For larger codebases, scope the scan to specific directories (auth, API routes, data layer) and rotate coverage across weeks.

Managed Agents Build 3: API migration agent

If you're deprecating an API and have dozens of services still calling it, this agent maps every call and writes the migration diffs.

You’re deprecating a v1 API. You have 47 services calling it. Nobody knows which endpoints each service uses.

This agent takes the old API spec and the new API spec. Reads every file across every service. Identifies every call to the old endpoints. Writes the migration diff for each service. Opens draft PRs.

You are an API migration agent.

Context:
- Old API base URL: [V1_API_URL]
- New API base URL: [V2_API_URL]
- API change summary: [PASTE DIFF OF BREAKING CHANGES]

When triggered:

1. Read every file in the repository looking for calls to
   the old API base URL or deprecated endpoint paths.

2. For each file with old API calls:
   - List every deprecated call found
   - Map each to its v2 equivalent based on the API change summary
   - Note any calls with no direct v2 equivalent (flag for manual review)

3. For files with clear 1:1 migrations only:
   - Write the updated file with v2 calls
   - Open a draft PR: "Migration: [filename] v1 → v2 API"

4. For files requiring manual review:
   - Create a Notion page: "API Migration — Manual Review Required"
   - List each ambiguous call with context

5. Post to Slack channel [CHANNEL ID]:
   "API migration complete. [X] files auto-migrated (draft PRs
   open). [Y] files need manual review: [Notion URL]"

Never merge draft PRs automatically. Human review required on every PR.

These three builds are engineering workflows, but the same architecture works for any domain: data quality checks across warehouse tables, content moderation pipelines, customer onboarding sequences, or compliance document review. The pattern is the same - define the agent, scope its tools, let it run.

5. What’s live, what’s preview, and what breaks

Public beta, available today: Routines on all paid plans (Pro 5/day, Max 15/day, Team and Enterprise 25/day). Managed Agents on the API - requires an API key, separate from your Claude chat subscription - with the built-in tool set (bash, file ops, web search, web browse, code exec, MCP).

Research preview, rate-limited: Webhook triggers on Routines. Multi-agent patterns where agents spawn sub-agents. Cross-session memory. Self-evaluation against success criteria.

Known limits that break builds:

• Routines are fully stateless between runs. No workaround except external storage.

• Webhook triggers don’t retry on failure. You need to handle at-most-once delivery on your end.

• Managed Agent sessions are capped at several hours. Long-horizon jobs need to checkpoint and resume.

• Compliance posture is still maturing. If you’re in HIPAA, PCI, or FedRAMP scope, check Anthropic’s trust center at trust.anthropic.com and confirm BAA/DPA availability before you design around this.

What to ship first

If you’re not technical: set up the competitor intelligence Routine or the weekly team digest. Takes an afternoon. Runs forever. Pro plan (5 runs/day) covers both. If you’re running 3+ daily Routines, upgrade to Max.

If you ship code: the PR reviewer and dependency audit are the fastest wins. Three Routines are running in my setup right now.

One thing nobody tells you: when a Routine fails, it fails silently. No Slack message, no email. Check your Routine’s run history at claude.ai/code/routines after the first scheduled run. If you see a red status, read the log, fix the prompt, and hit Run Now before re-enabling the schedule.

For Managed Agents, pick the build that addresses your biggest current bottleneck. Security coverage gaps: run the scanner weekly. Slow bug resolution: wire the bug-to-PR agent. API migration backlog: the migration agent clears it in one session.

Ship one. Review the output for four weeks. Then ship the next.

Final Words

The teams pulling ahead in the next year won’t be the ones with the most agents.

They’ll be the ones wiring these into permanent workflows, reviewing outputs weekly, and adjusting prompts the moment something drifts.

One Routine this afternoon. One Managed Agent this month. Review both for four weeks before you ship the next.

Subscribe now

PS. The Product Growth edition covers Routines and Managed Agents from the PM side. If that’s useful for your team, it’s at news.aakashg.com.

Beyond AI, here’s what I found interesting this week:

1/ Tennis players live 9.7 years longer than sedentary people in the Copenhagen City Heart Study, which followed 8,577 Danes for 25 years. Badminton added 6.2 years, soccer 4.7, cycling 3.7, swimming 3.4. The authors flag this as observational, so association not causation, but the ranking is striking: the sports with the most social interaction added the most years. You literally cannot play tennis alone. The forced social component is doing a lot of the work.

2/ France quietly pulled its gold out of New York and moved it to Paris. The official reason was “higher quality bars.” The real reason: they made €12.8 billion in profit on the swap and exited the U.S. financial system while calling it maintenance. Germany and Italy have $245 billion still sitting in Manhattan. Politicians in both countries are already asking for it back.

3/ Leopold Aschenbrenner sold every share of Nvidia last quarter and bought fuel cells, Bitcoin miners, and power companies instead. His logic: you can order 100,000 GPUs in six months. You cannot add 500 megawatts to the grid in six months. Oracle just signed a 2.8 GW fuel cell deal with his largest holding. The stock jumped 15% after hours. He’s 24.

4/ The flat stripe on your towel is called a dobby border. It’s the only part engineered to not absorb water. It exists to stop the edges from unraveling in the wash, control shrinkage, and give hotels a flat surface to embroider their logos on. The part that feels decorative is doing all the structural work.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job search coaching? Apply to my cohort.

Share

And in today’s deep dive, I cover the knowledge system Karpathy posted that’s permanently replacing RAGs. It’s the first thing I’ve tried that makes your AI knowledge compound instead of reset.

Subscribe now

Ariso: Your AI Chief of Staff

I had 80+ loose files rotting in my Google Drive root. Asked Ariso’s AI to organize them.

Twenty minutes later: GTM, Product, Finance, Legal, Content - all sorted, no rules or templates needed. It just read the files and figured it out. One less Saturday wasted on admin.

Try it free at ariso.ai/aakash.

Try Ariso

There’s a billion AI news articles every week. Here’s what actually mattered.

News

Anthropic just made every agent orchestration startup obsolete

Claude Managed Agents launched this week in public beta. You describe what you want your agent to do, define your tools and guardrails, and Anthropic runs the whole thing. Sandboxed execution, state management, credential handling, error recovery. All of it, handled.

What used to take months of infrastructure work now takes days. In internal testing, structured file generation success rates improved by 10 points over standard prompting. Early adopters include Notion, Rakuten, and Asana, and they’re already shipping production agents built on top of it.

Here’s what people are missing. Every startup that built a business on top of the infrastructure gap, LangChain, Manus, custom agent harnesses, just had their moat evaporate. Anthropic isn’t just offering a faster path to deployment. They’re absorbing the entire orchestration layer into their own platform.

The pricing tells the story. $0.08 per agent runtime hour, on top of model usage. That’s not an enterprise premium. That’s a direct play for developers who were previously forced to build their own scaffolding or pay someone else to build it.

Why this matters for you. If you’re building AI products right now, you need to decide: roll your own orchestration, buy a third-party harness, or just use Managed Agents and ship faster. The third option just got a lot more credible.

The way I see it, this is Anthropic pulling the rug out from under the middleware layer of the AI stack. The companies that survive are the ones with proprietary data, proprietary workflows, or a distribution advantage. Infrastructure players just lost a round.

Other News that Mattered this Week

• Google released Gemma 4 , the first time the Gemma family ships with a truly open commercial license. The 26B MoE model activates only 3.8B parameters at a time, and the edge models run on Raspberry Pi. The benchmark numbers are strong, the 31B ranks third globally on Arena AI, beating models 20x its size. But the license change is what actually matters. No usage restrictions buried in fine print, no legal review friction. Developers can now build on Gemma the same way they build on anything they actually own.

• Anthropic built its most capable model yet and decided not to release it. Claude Mythos Preview found a 27-year-old OpenBSD vulnerability, broke out of its sandbox during testing, and emailed a researcher to confirm it had done so. Instead of a public launch, Anthropic formed Project Glasswing: a restricted coalition of 40+ companies including Apple, Google, Microsoft, and Amazon with $100M in model credits to use Mythos exclusively for defensive security work. Responsible release or competitive gatekeeping? Probably both.

• Meta finally released the first model from their team led by Alexandr Wang, Muse Spark. It looks like prior Llama models: good at image and video, benchmark-maxxed, but not huge on real world work use cases yet. Data analysis and counting in images are amongst its top skills.

• Claude’s M365 connector is now available on every plan, including free. Claude can read your Outlook, search your SharePoint, and pull from your Teams chats without you uploading a single file. Microsoft Copilot charges $30/user/month for essentially the same access. Anthropic just made that a much harder sell.

Resources

• AutoAgent is the open-source library every agent builder needs to know about right now. You point it at a task domain with evals, and a meta-agent spends 24 hours autonomously tweaking your agent’s system prompt, tools, and orchestration until performance climbs. No human tuning. It hit #1 on SpreadsheetBench (96.5%) and the top GPT-5 score on TerminalBench (55.1%), beating every hand-engineered entry on both leaderboards. I’ve been watching this closely because it changes the question from “how do I tune my agent?” to “why am I tuning my agent at all?”

The AI Second Brain That Actually Works

Karpathy posted a knowledge base system that replaces RAGs and achieves what that Knowledge management system is not been able to do for years. I’ve been running it for a week and a half. It’s the first thing I’ve tried that actually makes your AI knowledge compound instead of reset.

That’s today’s deep dive:

1. Why every second brain you’ve built eventually died

2. What Karpathy actually figured out (and why it spread so fast)

3. Four use cases for engineers and builders

4. How to set it up in 30 minutes

5. What it looks like after 30 days

Subscribe now

1. Why Your Second Brain Keeps Dying

You hit a hard problem. You solved it. Three months later the same problem shows up in a different form and you can’t remember what you figured out the first time.

The folder exists. The notes exist. The solution is gone. It lived in your head and left when you moved on.

That’s not a discipline problem. Maintaining a knowledge base is a specific job: cross-referencing sources, updating summaries, flagging when new information contradicts old conclusions. You’re already running six things. The moment you fall behind by a week, the guilt of catching up is enough to make you never open it again.

Every AI tool has the same flaw underneath. NotebookLM, ChatGPT uploads, Notion AI: they all retrieve at query time instead of building anything persistent. Ask a question, get an answer, session ends. Tomorrow: scratch. Your knowledge never compounds. The AI never gets smarter about your domain.

The best engineers and builders don’t just know more. They forget less. And until now, no tool has actually helped with that.

2. What Karpathy Actually Figured Out

Karpathy hit the same wall with ML papers. Every session started from scratch. Ask a question connecting three papers, get the answer. Come back the next day: the AI had forgotten everything it built.

He tried every tool designed to fix this. Same result. So he asked a different question: what if instead of waiting until query time to process your sources, the AI compiled them once at ingest and maintained that compilation as new sources arrived?

He built that system. 18.7 million views on X. 5,000+ GitHub stars.

Three pieces make it work.

1. Raw is your junk drawer. Articles, transcripts, docs, notes. Dump it in without organizing. That’s the AI’s job.

2. Wiki is what the AI builds from raw. Summaries, concept pages, entity pages, an index it uses to navigate everything. You never write the wiki. The AI writes and maintains it entirely, every session.

3. Schema is a single CLAUDE.md file that tells the AI what your knowledge base is about and how to organize it. You write this once. The AI follows it every session.

The tedious part of maintaining a knowledge base is the bookkeeping. Humans abandon wikis because the maintenance burden grows faster than the value. The LLM doesn’t get bored. That’s the shift.

3. Four Use Cases

Use Case 1 - Stakeholder Memory

The wiki ends up knowing your CTO better than most new employees do after their first month.

Every meeting note, Slack thread, and email thread goes into raw/. The wiki builds a page per stakeholder: what they care about, what they’ve pushed back on, what language has landed, what they’ve approved and why.

Before your next meeting with the CTO, run this:

Pull the stakeholder page for [Name].
What have they pushed back on before?
What framing has landed with them?
What should I lead with tomorrow?

You walk in knowing exactly what objections are coming, what framing will land, and what to address defensively. Most PMs figure this out live, in the meeting. Often too late.

The relationship context you were never going to write down consistently is now written down. Every meeting. Every decision. Every signal. All queryable.

Use Case 2 - Side project context that doesn’t evaporate

You work on your main job five days a week. Your side project gets weekends. By the time you open it Saturday morning you can’t remember what you were building toward. You spend the first hour just getting back up to speed.

Drop your planning notes, decisions, and TODOs into raw/ after every session. Write a brief note before you close the laptop about where you left off. The wiki maintains your running context.

Saturday morning prompt:

Where did I leave off on [project]? 
What was I trying to solve and what's the next step?

You start building in 5 minutes instead of 60.

Use Case 3 - Team Onboarding That Doesn’t Lose Knowledge When People Leave

A senior employees leaves your team. Years of context walk out with them. Why you made that architecture call 18 months ago. What users said about the feature you cut. What the CTO actually cares about versus what they say they care about.

If the wiki exists, the transition is painful but survivable. If it doesn’t, the next employees starts over.

Drop past docs and meeting notes into raw/. A new employee joins, they read the wiki, they get 80% of the context you’d otherwise spend weeks transferring. Karpathy named this use case directly in his gist: “Business/team: an internal wiki maintained by LLMs, fed by Slack threads, meeting transcripts, project documents, customer calls. The wiki stays current because the LLM does the maintenance that no one on the team wants to do.”

That’s your second brain working at team scale. The institutional memory that usually walks out the door when someone leaves, finally written down somewhere the next person can actually find it.

Use Case 4 - Solutions you solved once and forgot

You spent four hours debugging a gnarly issue last month. You found the fix. You moved on. Now it’s back, in a slightly different form, and you’re starting from zero.

Drop your debugging notes into raw/ when you solve something non-obvious. The wiki builds a page per problem type. Next time it surfaces, you query before you debug.

Have I hit anything like [error / behavior] before? 
What was the fix and what caused it?

The answer is usually there. The problem is you never had anywhere to put it.

4. How to Set It Up

Path A: Use My Skill

Use this skill I’ve created for you:

Grab the Skill

Path B: Do it Manually

Create the folder structure:

mkdir -p my-knowledge-base/{raw/assets,wiki,outputs}
cd my-knowledge-base
git init && git add . && git commit -m "setup"

Four folders. Raw is source material the AI reads but never modifies. Wiki is what the AI builds. Outputs is where answers to queries live. Git gives you version history for free.

Write a CLAUDE.md that tells Claude what this knowledge base is for, how to categorize sources, and what to do when new sources contradict old ones. The key conventions: every wiki page cites its source, cross-references use [[page-name]] links, and an append-only log tracks every ingest.

Start ingesting one source at a time:

Read the schema in CLAUDE.md. Process [FILENAME] from raw/. 
Read it fully, discuss key takeaways with me, then: create a 
summary page in wiki/, update wiki/index.md, update all relevant 
pages, flag any contradictions with existing wiki content, 
and append an entry to wiki/log.md.

Read the wiki summary after each ingest. Ask one question. Do that five times and you already have 15-30 interconnected pages.

One thing to resist: the instinct to dump everything at once. Batch ingesting produces worse results. You lose the ability to guide what the AI emphasizes. Karpathy ingests one source at a time. So do I.

5. What 30 Days Looks Like

By week three, something shifts. You ask a question and the wiki answers with something you forgot you knew. A note from two months ago surfaces in response to a problem you’re facing today. A contradiction between two things you read six weeks apart gets flagged automatically.

That’s the compounding effect. It doesn’t happen in one impressive moment. It accumulates quietly, source by source.

After 30 days you have 50-80 interconnected pages built from your own work. Live pages per project, per library, per problem type. A growing collection of answered questions that represent your best thinking on the things you keep coming back to.

The wiki made me a better researcher because knowing it would compound made me more deliberate about what I put in. The raw folder is the cleanest set of notes I’ve ever kept.

Here’s where I land: the best builders don’t just know more. They forget less. This is the first system I’ve used that actually delivers on that.

That’s it for today’s deep dive. I wrote a full version of this for PMs in Product Growth. It includes a complete CLAUDE.md starter template you can drop straight into your project, and a setup skill which helps you the complete second brain architecture in just 60 seconds.

Beyond AI, here’s what I found interesting this week:

1/ Peanuts in Coke is one of the most accidentally perfect food pairings ever. The acid releases glutamate from the peanuts, generating umami in real time. The salt makes the sweetness hit harder. Georgia farmers invented this in the 1920s because they needed one hand free while working. They nailed the optimal flavor loop a century before food science could explain why.

2/ Apple accidentally built the world’s largest hearing aid company. AirPods Pro 2 got FDA clearance as a clinical-grade hearing aid in 2024. Prescription devices cost $4,700. AirPods cost $249. Every hearing aid company spent decades engineering invisible devices to reduce stigma. Apple made earbuds a status symbol first, then added the medical function as a software update.

3/ Nestlé’s response to the KitKat truck heist is the best unplanned marketing campaign of the year. 413,793 bars stolen somewhere between Italy and Poland. Still missing. Their move: fake presidential motorcade through Canadian cities. Security guards at retail displays. A job posting requiring “experience guarding high-value assets.” The stolen cargo is worth $620K. The Dexerto post hit 4.7M views. The thieves wrote the creative brief for free.

4/ Your morning coffee on an empty stomach is a cortisol bomb. Cortisol peaks when you wake up. Breakfast is the signal that pulls it back down. Skip it, add caffeine, and you’ve stacked two stress drivers before 9 AM. A meta-analysis across 399,550 people linked skipping breakfast to higher odds of depression. Your body uses that first meal as a hormonal reset. Skip it and your stress system runs in the background all morning like an app you forgot to close.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job searching coaching? Apply to my cohort.

Share

And in today’s deep dive I cover the 120+ features Anthropic has shipped in the last 90 days.

I rank them into S, A, B, C, and D tiers. Then I give you the key details on the S tier.

Subscribe now

Ariso: Your AI Chief of Staff

I am using an under-the-radar AI tool that talks to me in Slack - and it’s not OpenClaw. It’s Ariso:

Ariso connects to my email, calendar, and docs, then handles meeting prep, follow-ups, and the rest. It remembers every conversation and gets smarter the more I use it.

Built by cybersecurity founders from Durham NC (where I used to live). I personally recommend it. Best of all? It’s free right now: ariso.ai/aakash

Try Ariso

There’s a billion AI news articles every week. Here’s what actually mattered.

The Week’s Top News: Anthropic Leaked Claude Code’s Entire Source Code

The “safety-first AI lab” just had two major leaks this week: first, that its new Mythos model is so good it poses cybersecurity risks; and, second, ironically, the source code for Claude Code.

A source map file shipped inside the Claude Code npm package by accident. Chaofan Shou spotted it first and posted the download link on X.

Within hours, 512,000 lines of TypeScript were mirrored across GitHub, forked 41,500+ times, and picked apart by thousands of developers. Everyone is fixating on the embarrassment. But the real story is what’s inside the code.

I went through it, and there’s a lot. 44 feature flags for capabilities that are fully built but haven’t shipped yet. A persistent background agent called KAIROS that keeps working while you’re idle. An anti-distillation system that injects fake tool definitions to poison competitors trying to scrape Claude Code’s API traffic. And an “undercover mode” that strips all Anthropic references from commits when employees contribute to public repos.

Anthropic is preparing for an IPO, selling enterprise security, and positioning as the responsible AI company. Two leaks in one week make that pitch a lot harder.

The Other News That Mattered

News

• Sam Altman’s non-binding DRAM deals may have caused the worst consumer hardware crisis in a decade. He locked up 40% of global memory supply with letters of intent that were never real. DDR5 prices jumped 171%. Micron killed its 29-year-old Crucial brand to chase AI demand. MU is now down 33% from its post-earnings high. The signature was worth exactly what the paper said: nothing binding.

• Meta quietly released TRIBE v2, a model that predicts how your brain responds to anything you see, hear, or read. 70x higher resolution than v1. They open-sourced it, which tells you the research advantage is already captured. A company spending $135B in capex didn’t build this for academic citations.

Resources

• Jack Dorsey published From Hierarchy to Intelligence. The thesis: corporate hierarchy exists to route information, and AI can now do that continuously at scale. Block cut 4,000 jobs and framed it as acceleration, not cost reduction. Fascinating read. Whether you agree with Dorsey or not, every leader should read this and decide what it means for their org.

New Tools

• Google AI Quests is expanding to eight new languages. It’s a free, gamified AI literacy experience built with Stanford that teaches students aged 11-14 how AI actually works through real research scenarios like flood forecasting and disease detection. If you have kids or work in edtech, worth bookmarking.

Fundraising

• OpenAI closed $122 billion at an $852B valuation, the largest private funding round in history. But look at who wrote the checks: Amazon ($50B), Nvidia ($30B), SoftBank ($30B). All three are also OpenAI’s biggest vendors. The capital isn’t just flowing in, it’s circulating. $35B of Amazon’s piece is contingent on an IPO or AGI by 2028.

• WHOOP raised $575M at a $10.1B valuation. A screenless rubber band with SaaS metrics: 2.5M members, $1.1B bookings run rate, 103% YoY growth, cash flow positive. Qatar Investment Authority, Mubadala, Abbott, and Mayo Clinic all wrote checks. Sovereign wealth funds and healthcare institutions don’t invest $575M because recovery scores are cool. They’re buying a distribution channel for continuous patient monitoring outside the clinic.

I Ranked Every Feature Anthropic Shipped This Quarter

Anthropic shipped 120+ features in 90 days.

Most people saw a few headlines and moved on. Agent Teams, Opus 4.6, maybe Dispatch. But they’re missing the bigger picture. These features aren’t isolated launches. They build on each other in ways that aren’t obvious if you’re only watching one product line.

I researched every Anthropic release so far this year. Blog posts, changelogs, tweets from the engineers who built each feature, official docs, community reports. And I tested them.

Honestly, most of it is incremental. But about 30 features are worth your time.

Here’s what you need to know:

The Full Picture

I mapped every release to a calendar. Here’s what Q1 looked like for Claude:

That’s 40+ Claude Code releases, 15+ Cowork updates, 20+ API changes, and two new models in 90 days.

The thing that stands out when you see it on a calendar is the overlap. Agent Teams and Opus 4.6 shipped the same week. Code Review landed on a Monday, and by Friday they’d added 1M context GA and four more Claude Code features. No product line was waiting for another.

That matters because these features build on each other. /loop is more useful when the model can stay on task for 14 hours. Dispatch gets more powerful when it can launch Claude Code sessions. Worktrees work better when 1M context means each agent can hold the full repo. The product lines connect in ways that aren’t obvious if you’re only watching one.

I went through all of it. Here’s the tier list:

Now let’s cover everything in S tier.

S Tier: Claude Code

Auto Mode

Auto Mode lets Claude make permission decisions on your behalf. It runs two safety checks before every action.

I enabled it and gave Claude a refactor task: dependency injection across a user service, update all tests, make sure CI passes. Claude edited 14 files, ran tests 3 times, fixed two test failures, and committed. Zero permission prompts. 8 minutes total. Before Auto Mode, I would have clicked “Allow” roughly 40 times during the same task.

claude --enable-auto-mode

Not the same as --dangerously-skip-permissions. Auto Mode still won’t delete your database. It just stops asking permission to edit a CSS file.

Opus 4.6

Opus 4.6 launched in February and the world hasn’t remained the same ever since.

Everyone focused on the benchmarks. The part that matters is the task horizon. Opus 4.6 can stay on a single task for over 14 hours without losing track. That’s what makes /loop, Agent Teams, and Dispatch viable on real projects.

Without this model, half the features on this list break on anything complex.

Agent Teams

Agent Teams lets a lead agent delegate to up to 10 parallel teammates that coordinate through a shared task list.

I tested this on a side project with three open issues: an auth flow, a dashboard layout, and a broken CSV export. Described all three to the lead agent and let it delegate. All three had PRs open within 40 minutes. Two passed tests immediately. The third needed one manual fix on a type mismatch.

Add CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 to your environment and describe multiple tasks in a single prompt. Requires Opus 4.6. Research preview.

Cowork Launch

Cowork is Claude running persistently on your computer with access to your apps, files, and screen. You message it, walk away, and come back to finished work. It’s a great skin on Claude Code:

Before a 2-hour meeting, I told Cowork: “Catch up on my Slack DMs and summarize anything I need to respond to. Then find the Q1 metrics deck in my Drive and update the revenue chart with this month’s numbers.”

When I got back, Cowork had summarized 12 Slack DMs into three priority buckets, flagged two that needed immediate replies, and updated the chart. The Slack summary saved me 30 minutes of scrolling.

Connectors

Connectors let you plug Claude into your work tools with one click. Gmail, Slack, Notion, Figma, Asana, Google Drive, and 50+ others as of February 2026. Built on MCP, but you don’t need to know that. You just toggle it on.

I connected the Figma connector and pointed Claude Code at a mockup for a settings page. It read the frame, extracted the layout and tokens, and generated a React component that matched the design. When I asked it to adjust spacing, it updated both the code and pushed the changes back to the Figma file. Design-to-code and code-to-design in one loop.

Connectors work across Claude.ai, Cowork, and mobile. Any MCP server anyone builds becomes a connector you can add. Refer docs here.

/loop

/loop runs a task on a repeating interval inside your session. These are the patterns I’ve been using:

/loop 5m /babysit
/loop 30m /slack-feedback
/loop 1h /pr-pruner

The first auto-addresses code review comments and pushes PRs to production every 5 minutes. The second puts up PRs for Slack feedback every 30 minutes. The third closes stale PRs hourly. These patterns come from the Claude Code team’s own workflow.

Loops run for up to 7 days and support up to 50 tasks per session. /loop can invoke other slash commands, so you can chain it to a skill and the whole workflow runs without you touching the terminal. I covered the origin of this pattern in Ship Features While You Sleep.

Final Words

Start with three things today. Turn on Auto Mode so you stop clicking “Allow” forty times per session. Set up Cowork so Claude works while you’re in meetings. And run /loop 5m /babysit on your next PR.

I’m shipping faster and I’m not working more hours. Not every feature here will matter for your workflow, but the S tier ones changed mine.

That’s it for today’s deep dive.

I wrote an extended version for paid Product Growth Subscribers covering A, B, C, and D Tier features. Read it on Product Growth →

Beyond AI, here’s what I found interesting this week:

1/ Zuckerberg’s yacht once sailed 9,600 nautical miles to the South Pacific and waited two months for him to show up, but he never boarded and the vessel turned around empty. It burns 1,165 gallons of diesel per hour and in nine months produced what 400 American households emit in an entire year. Meta pledges net zero across their value chain by 2030. Net zero is a line item in someone else’s budget.

2/ The S&P 500 had its worst quarter since 2022. Forward P/E fell to 19.7x, below the 5-year average. But every time it’s dropped below 20x since 2020, forward returns have been positive. Miss just the 10 best trading days and your returns get cut in half. The worst time to look at your portfolio is the best time to add to it.

3/ GTA 6 isn’t delayed because $3B is hard to spend. GTA 5 still generates $700 million a year thirteen years after launch, and day one of GTA 6 is also the last day of GTA Online’s current economy. Every delay is Rockstar protecting the most profitable entertainment product ever made while making sure its successor can match that run rate on arrival.

That’s all for today. See you next week,

Aakash

P.S. Want my AI tool stack? Join my bundle. Want my job searching coaching? Apply to my cohort.

Share

Everyone is comparing it to OpenClaw and Claude Cowork. But Computer is doing something different.

I’ve been testing it for weeks. Today’s deep dive is your complete guide to it.

Ariso: Your AI Chief of Staff

I had 80+ loose files rotting in my Google Drive root. Asked Ariso’s AI to organize them.

Twenty minutes later: GTM, Product, Finance, Legal, Content - all sorted, no rules or templates needed. It just read the files and figured it out. One less Saturday wasted on admin.

Try it free at ariso.ai/aakash.

Try Ariso

There’s a billion AI news articles every week. Here’s what actually mattered.

The Week’s Top News: OpenAI has left the video generation business

OpenAI just exited the video generation business entirely. App dead. API dead. No video inside ChatGPT. Disney’s $1 billion deal, signed four months ago, is dead.

Read that again. This isn’t a consolidation into the super app. Altman told staff Tuesday that OpenAI is winding down all products using video models. Disney’s own statement says they respect OpenAI’s decision to “exit the video generation business.” The Sora research team is being redirected to robotics.

The reason is sitting right there in the competitive data. Anthropic hit $19 billion in annualized revenue by early 2026 selling text and code. No video generation. No image generation. No consumer social app. No Disney deal. One product surface: chat, code, computer use, all in one place. OpenAI looked at where every dollar of market growth was coming from and saw the answer: coding and enterprise.

So now they’re copying the model. ChatGPT, Codex, and the browser merge into one app. Instant Checkout killed today too. Every consumer experiment is getting cut. What remains is the Anthropic playbook: one app, code and chat, enterprise and developer focus.

The Sora numbers explain the urgency. Total consumer revenue across iOS and Android since September: $1.4 million. Peak month was $540,000. Every video generation burned GPU compute that could have been running inference for ChatGPT or Codex instead. OpenAI’s own head of Sora announced generation limits because chips couldn’t keep up. At $14 billion in projected 2026 losses, every GPU matters.

Google just inherited the AI video market by default. Nano Banana already lives inside Gemini. No standalone app to manage, no separate brand to support. Among the majors, they’re the only ones left. Runway, Kling, Minimax, Luma, and the other independents are still shipping, but none of them have Google’s distribution.

Disney put $1 billion in stock warrants on a product that lasted six months. The deal was announced in December. Characters from Marvel, Pixar, and Star Wars were supposed to be generating fan videos on Sora by now. Instead, Disney is writing a polite press statement about “respecting OpenAI’s decision” while its legal team unwinds a deal that never produced a single licensed video.

Four months from billion-dollar partnership to obituary. That’s how fast the AI product landscape reprices when the unit economics don’t work.

The Other News That Mattered

• Anthropic shipped four features in 30 days that reverse-engineer every OpenClaw capability into an enterprise-safe product. Dispatch, Claude Code Channels, computer use in Cowork, curated plugin marketplace. The agent race rewards shipping velocity over hiring velocity.

• Someone poisoned LiteLLM, the Python package that routes AI API keys for NASA, Netflix, and Stripe. 97 million downloads a month. A single pip install exfiltrated every SSH key, cloud token, and .env file on the machine. The only reason it got caught is the attacker vibe coded the malware so badly it crashed computers before it finished. If you’re shipping AI agents, your supply chain is your attack surface.

• Apple put AI coding agents in Xcode 26.3, then blocked Replit and Vibecode from updating in the App Store. Replit dropped from first to third in developer tools downloads in two months. Every app built through vibe coding that ships as a web app is revenue Apple never touches. The policy isn’t about code execution rules from 2009. It’s about who gets to be the on-ramp.

• Google Research published TurboQuant, a compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup on H100 GPUs. The internet is already calling it Pied Piper. Cloudflare’s CEO called it Google’s DeepSeek moment. The paper will be presented at ICLR 2026 next month.

Resources

• If you’re using Claude Code without plan files, you’re doing it wrong. Matt Van Horn’s piece on every Claude Code hack he knows is the best single resource I’ve seen this month. I have 70 plan files and 263 commits in the last 30 days. Plan first, build second.

Tools

• Claude Code cloud scheduling is the feature I've been waiting for. Set a repo, a prompt, a cadence. Your laptop can be off. Nightly dependency audits, stale PR reviews, test reruns, all running at 3am on Anthropic's infra. $200/month.

• MCP Apps hit mobile this week. Figma, Canva, Amplitude, all inside Claude on your phone. You stop opening apps. You open Claude and Claude opens them for you. The race to own the AI routing layer on your phone is now a three-way fight.

• Dispatch lets you run multiple Cowork sessions from your phone in Claude Code. Spawn a research task, a draft, and a gap analysis from the same thread while you're walking your dog. Each task runs independently on your desktop. Your phone is the command chair.

I'm doing a free masterclass on April 7th covering the two interview rounds that eliminate most PM candidates, behavioral and AI PM, and exactly how to pass them. Register here.

Register here

The Complete Guide to Perplexity Computer

The agentic AI world is moving fast. First it was Manus, Openclaw, Claude Code, then Claude Cowork. Now Perplexity has dropped Perplexity Computer.

They all want to be the tool you can’t live without. But how do you know which one to pick for your work? And how do they actually compare when you put them to work?

I have a feeling Computer is about to earn the same loyalty.

That’s what this guide is about. Perplexity Computer treats vague instructions as permission to do far more than you asked. And that costs real money. I spent my first week learning that the hard way.

But once I figured out how the system actually works, I built a prompting structure that controls the cost before Computer starts spending. And I built workflows that genuinely replaced hours of manual work.

I’m not here to inflate AI hype. I’m here to share what actually works.

Here’s what you need to know:

1. What Is Perplexity Computer

2. How to Use Perplexity Computer

3. Honest Limitations and Structural Risk

Subscribe now

1. What Is Perplexity Computer

Perplexity Computer launched on February 25, 2026. CEO Aravind Srinivas calls it a “general-purpose digital worker.”

Despite its name, it’s not a physical computer. It’s a cloud-based system made up of multiple AI agents working together.

It can research, design, build, test, deploy, and automate. With 400+ app integrations, it can also read your Gmail, push code to GitHub, query your Snowflake database, post to WordPress, and about 396 other things I’m still discovering.

All from a single prompt and a single interface. No terminal. No API keys. No local setup.

That last part matters. If you’ve tried OpenClaw , you know the headache. Local installations, environment configs, dependency hell. Computer runs entirely in the cloud. You describe the outcome. Computer handles everything else.

The 19 Models Under the Hood

I was jumping between Claude, Gemini, and GPT constantly. Different tabs, different contexts, different formatting. Copy-pasting outputs from one model into another just to get a complete workflow done. It was exhausting.

Computer eliminates that entirely. It sits on top of 19+ frontier AI models. Claude Opus 4.6 handles the core reasoning. Gemini runs deep research tasks. ChatGPT 5.3 assists with long-context recall. Grok takes the lightweight, quick tasks. Nano Banana generates images. Veo 3.1 produces videos (Google’s models).

Computer picks the best combo for the job. You write one prompt. Computer routes the work.

Tip: You can manually assign which models handle which parts of your task. Most people skip this and let Computer pick the most expensive option. Learning to control model routing is the fastest way to cut costs.

The Business Model Shift

Perplexity spent three years building a search engine at $20/month flat. Computer layers a consumption-based product on top of that entire base. Max subscribers get 10,000 credits per month. Usage-based pricing. Spending caps. You choose which models run your sub-agents.

This is a fundamentally different business than search. Search is a flat subscription. Computer is metered compute. That changes how you use it, how you budget for it, and how you think about what’s worth automating.

How Tasks Actually Run

Everything in Computer is organized around Tasks. A Task is not a conversation. It’s a job.

When you submit a prompt, the orchestrator (powered by Claude Opus 4.6) reads it, breaks the objective into subtasks, creates a separate sub-agent for each one, assigns each to the best-suited AI model, runs them in parallel, and compiles results in the right panel.

Every sub-agent consumes credits. More sub-agents means higher cost. This is why vague prompts are expensive. The orchestrator creates sub-agents for every plausible interpretation. Specific prompts create fewer, more targeted agents.

You can see the decomposition happening in the output panel as it runs.

Tasks run asynchronously in the cloud. Close your browser. Go to dinner. When you come back, everything is done.

Scheduled tasks take this further. Your machine doesn’t need to be on. Set it up once. Computer runs it on a recurring schedule whether your laptop is open, closed, or sitting in a drawer. This kills the need for OpenClaw to be honest, as Computer has zero local setup.

The Sandbox

Every task runs inside an isolated Firecracker microVM (same tech AWS built for Lambda). Real filesystem, real browser, real tool integrations. Complete isolation between tasks.

Computer can’t touch your local files. Everything runs in the cloud. If a task goes wrong, the failure stays contained. That’s the safety advantage over OpenClaw, which runs locally with direct access to your filesystem.

The sandbox also has advanced anti-blocking architecture. It navigates around anti-bot blockers that normally stop AI agents from browsing the web. That’s why it can scrape competitor sites and interact with web apps other agents can’t.

Limitations: no desktop app control, no screen/mouse access, file access limited to uploads and connectors.

Setting Up

You need a Perplexity Pro subscription ($20/month), Max ($200/month), or Enterprise Max ($325/seat/month). Pro gets 4,000 credits per month. Max gets 10,000. At launch, Max users received a one-time bonus of 20,000-35,000 extra credits (valid for 30 days).

Getting started takes 30 seconds. Go to perplexity.ai, log in, click the Computer icon in the left sidebar. You’re in. No local setup. No terminal. No environment variables.

Credit counter in the sidebar updates in real time. Usage button in the top right of any task shows its specific cost. Full breakdown at perplexity.ai/account/usage.

Connectors (And Why They’re the Real Product)

This is where Computer stops being a research tool and becomes a workflow execution system.

Connectors let Computer access your actual data and take real actions in your existing software. Not summaries. Real read-and-write access.

Setting them up takes about a minute per app. Click Connectors in the sidebar, browse the list, click Enable, complete OAuth. Done. Computer can now read from and write to that service.

Here’s what makes Computer’s connectors different from other AI tools. Most agents give you 10-20 integrations focused on productivity basics. Computer has 400+, and the interesting ones go way beyond Slack and Google Drive.

What you can actually do with these connectors.

Here’s what surprised me about these connectors. They’re not just productivity integrations. Some give you access to things that normally cost thousands or require engineering to set up.

• I asked Computer to build me an investment memo on Shopify. It pulled live financials from FactSet, earnings transcripts from Quartr, analyst sentiment from finance Twitter. No API keys. No data subscriptions. Just built-in access to FactSet, S&P Global, Coinbase, CBOE, Nasdaq, NYSE. All native.

• Through Plaid, you can link your actual brokerage and Computer builds you a live portfolio dashboard. Daily P&L, position cards, news feed, earnings calendar, price alerts. A personal Bloomberg terminal on a private URL. Full setup and prompt in Use Case 4.

• Research access is wild. CB Insights, PitchBook, Statista. Computer bypasses the paywalls. A $5,000 market sizing report? Computer just pulls it.

• The Snowflake connector made me stop and think. Connect it, and Perplexity auto-generates a Data Map of your schemas. Type “What were the top 10 customers by revenue last quarter?” Computer writes the SQL, runs it, hands you charts. Plain English in, data out. Same with Salesforce, HubSpot, Databricks.

• GitHub lets Computer write, refactor, and push code to your repos. Gmail, Slack, LinkedIn go deeper than you’d expect. I had it scan my inbox, summarize my newsletters, draft unsubscribe requests, and calculate my SaaS burn rate from receipt emails. Tag Computer in a Slack channel and it runs research without you switching tabs.

• Polymarket was unexpected. Live prediction market odds for industry events. Real-time sentiment without refreshing a dozen tabs. Reddit for early business signals. Toggl Track for time tracking. And this caught my eye: direct purchases through PayPal. An AI agent that can buy things. That’s new.

• For SEO, Ahrefs, Google Search Console, Google Analytics, WordPress, Cloudflare. Full audit, competitor analysis, keyword mapping, content drafting, and live site updates. One prompt.

• And the one that surprised me most. Computer connects to Amazon’s Selling Partner API and Shopify natively. If you’re running D2C, that changes everything.

Bring your own tools. If Perplexity doesn’t have a connector for something you use, provide an MCP (Model Context Protocol) server URL. This lets you hook Computer up to proprietary CRMs, custom analytics servers, or private APIs. Enterprise admins can share custom connectors across the organization.

I wrote about how Computer replaced $225K/year in marketing tools. There are 15,384 MarTech tools on the market right now, and the average marketing team uses only 33% of their stack’s capabilities. Companies are paying more for tools they use less every year. That’s the gap Computer is attacking. The argument is that the entire category is waste.

Skills and Custom Instructions

A Skill is a saved set of instructions that auto-activates when Computer recognizes a matching task. Same concept as Claude Skills or Cowork plugins.

Without Skills, you’re re-explaining your brand guidelines, formatting preferences, and reporting structure on every task. With Skills, you explain it once.

Computer ships with built-in Skills for Slides, Research, Research Report, and Chart. To create a custom one, click Skills in the sidebar, click + Create skill, and upload a .md file.

Skills stack. A research Skill can hand off to a report-formatting Skill, which hands off to a slides Skill. One prompt, full pipeline.

Tip: If you’ve explained the same thing to Computer twice, it should be a Skill. Brand guidelines. Formatting rules. Research methodology. Build it once, never repeat yourself.

Custom Instructions are different from Skills. Skills activate for specific task types. Custom Instructions apply to ALL tasks, all the time.

Keep them under 1,500 characters. The single most effective one I’ve found:

“Always come back to me and clarify any misunderstandings and challenge my thinking to make sure you’re very clear on the stated outcome, and let’s create a brief short plan before we build anything.”

That forces Computer to verify what you want before spending credits. One line. Massive savings.

The Perplexity Stack: Use the Right Tier

This is the part most people miss. Perplexity has four product tiers. Using Computer for a question Pro Search handles in 10 seconds is like hiring a full project team to send one email.

Perplexity Comet is the AI-native browser. Unlimited on Comet Pro. Tab-level browsing, page summaries, routine web research. 80% of daily AI work should happen here.

Pro Search answers factual questions quickly from a limited source pool. If the answer exists and you just need it collected, start here.

Deep Research synthesizes insights across 50+ sources into long-form reports. 20 runs per month on Pro. One run equals one serious research task.

Perplexity Computer is for multi-step projects that require parallel agents, tool integrations, code execution, and structured deliverables.

2. How to Use Perplexity Computer

You do not need to know how to code, manage API keys, or use a terminal. You write your stated intention in plain English. Computer autonomously researches the data, writes the code, and deploys a fully interactive web application with a live, shareable link.

Use Case 1: The Lead Machine

Instead of manually sourcing leads, Computer acts as a full-time SDR.

The workflow: For cold outbound, give it a list of 30 target companies. It researches each one, identifies the right decision-maker (bypassing the CEO to find the actual department head), and drafts 30 personalized emails referencing specific company milestones. It schedules follow-up sequences, leaving everything in your Gmail drafts for one final review.
Cold outbound prompt:

Here's a list of 5 companies I want to sell to: [LIST]. For each 
one, find the founder or relevant department head on LinkedIn (not 
just the CEO, find the actual decision-maker for [YOUR CATEGORY]). 
Research their company's recent news, funding, and pain points. 
Draft a hyper-personalized cold email for each one referencing 
something specific about them. Each email must reference at least 
one specific company detail. Schedule 3-day and 7-day follow-up 
sequences. Leave everything in my Gmail drafts for final review. 
Do not send any emails without my approval. Show me the company 
profiles and identified contacts before you start drafting.

PS: If any of the connectors are missing, computer will ask you to connect on the fly itself, making the configuration super easy.

View Full Task On Computer

Use Case 2: “Senior Advisor” Contract Auditing

Legal review means expensive billable hours. Computer catches blind spots before the lawyers even see the document.

Upload a partnership agreement or proposal. Computer reviews it like a senior consultant. It fact-checks claims against live web data. If a proposal claims a market grew by 43%, Computer checks whether it actually grew by 4%. It flags risky language and delivers a marked-up Word document with track changes.

Review this uploaded partnership proposal. Fact-check everything 
against live web data, identify deal terms, flag any vague or risky 
language, suggest improvements, and give me a top 5 concerns 
summary. Focus on financial terms, liability clauses, termination 
conditions, and any factual claims about market size or growth 
rates. No images, video, or code. Deliver it all in a Word document 
with track changes and comments. Before you start, propose your 
review approach and wait for confirmation. Flag any claims you 
cannot verify immediately.

Use Case 3: The Silent 8 AM Competitor Ping

Set up a recurring workflow that checks five competitors’ websites and X accounts every morning at 8:00 AM. It watches for pricing changes, feature drops, stealth launches, and brand mentions.

If nothing changed, it stays completely silent. No “all clear” messages. It only pings you when something strategic actually shifted.

Every morning at 8:00 AM EST, check these five competitors' 
websites for pricing changes, new feature announcements, and any 
new blog posts: [COMP 1], [COMP 2], [COMP 3], [COMP 4], [COMP 5]. 
Also search X for any mentions of them. If anything has changed 
since yesterday, send me a push notification with a summary. If 
nothing has changed, stay quiet. No daily summaries. No "all clear" 
messages. Maximum 5 items per report. Under 200 words. Only notify 
me if at least 2 qualifying changes are found.

View Full task on Computer

Use Case 4: Your Personal Bloomberg Terminal

This is one of the most impressive things I’ve seen Computer do. And you don’t need to write a single line of code.

The idea is simple. You connect your actual brokerage account, and Computer builds you a personalized finance terminal that runs on a private URL, updated automatically on whatever schedule you set.

Here’s the full flow.

Step 1: Connect your brokerage. Link Robinhood, Schwab, Fidelity, or whatever you use through Computer’s Plaid connector. Just say “connect my brokerage” and Computer walks you through OAuth. It never sees your login credentials directly. Once connected, it reads your holdings, positions, and transaction history.

Step 2: Build the terminal. Tell it “Build me a live portfolio dashboard that tracks my holdings.” Computer pulls your real portfolio data from the brokerage, fetches live market data using its native finance tools (FactSet, SEC filings, Coinbase, Quartr, all without API keys), and builds an interactive web dashboard deployed as a website.

Step 3: Make it always-on. The dashboard gets deployed with a private URL you can bookmark. Computer sets up recurring tasks to refresh data hourly, check for price alerts, earnings surprises, or significant moves. It sends you notifications when something noteworthy happens. A stock drops 5%. An earnings report lands. A position crosses your threshold.

Three things you need to do.

1. Connect a brokerage. Say “connect my brokerage.”

2. Describe what you want. “Build me a portfolio dashboard with my holdings, daily P&L, and news for each stock.”

3. Set up monitoring. “Check my portfolio every hour and notify me if any position moves more than 3%.”

No code. Computer handles data fetching, site building, deployment, and scheduling.

The build prompt:

Build me a live portfolio dashboard that tracks my connected 
brokerage holdings. Include current value, daily P&L, allocation 
breakdown, individual position cards with cost basis and gain/loss, 
a news feed for stocks I hold, an earnings calendar for my 
holdings, and configurable price alerts. Deploy as a private website 
I can bookmark. Do not make it publicly accessible. Dashboard only, 
no trading recommendations or financial advice. Before you build 
anything, propose the layout and wait for my confirmation. Show me 
a working preview before final deployment. Do not make it public 
without my approval.

Then follow up to make it always-on then give command to the task:

Set up a recurring task that refreshes my portfolio dashboard data 
every hour. Send me a push notification if any single position 
moves more than 3% in either direction, or if an earnings report 
drops for any stock I hold. If nothing noteworthy happened, stay 
quiet. Include the position name, change amount, and one line of 
context in each notification.

Use Case 5: Single-Ticker Investment Memos

If you’re exploring an acquisition, evaluating a stock, or preparing for a board conversation, this workflow replaces hours of manual research.

You can prep for a conversation about a potential investment. Would usually take an entire afternoon pulling data from 6 different sources. Computer hands a finished memo with charts in about 15 minutes.

View Complete Task on Computer

Build me an investment research memo on [TICKER SYMBOL]. Pull their 
latest financials and earnings transcript highlights. Compare their 
margins and growth to [COMPETITOR 1] and [COMPETITOR 2]. Check what 
analysts and finance Twitter are saying. Compile into a polished 
PDF with charts. I want a bull case, a bear case, and your final 
assessment. Maximum 10 financial sources. No video or audio. 3,000 
words max. After gathering the financials, show me the data summary 
before you start writing. If the same error happens twice, stop 
and ask me.

Use Case 6: Self-Auditing Your Marketing and SEO

Computer can turn the audit lens inward on your own assets.

Build a Skill defining great marketing for your brand. Computer crawls your website and scores every page. Instant visibility into what needs rewrites.

For SEO, give it your business details and let it run a full audit: Google Search Console data, competitor analysis, keyword mapping, content drafting, and technical fixes. All from one prompt.

Run a comprehensive SEO audit of [YOUR WEBSITE URL]. Compare to 
the top 3 competitors for [YOUR PRIMARY KEYWORDS]. Identify content 
gaps and draft a prioritized action plan. Use my connected Google 
Search Console and Ahrefs data. Focus on technical issues, content 
gaps, and keyword opportunities. Give me the action plan in order: 
technical fixes first, then content opportunities, then keyword 
targets. 3,000 words max. Do not make any changes to the website 
without my explicit approval. Before you start, propose the audit 
scope and wait for confirmation. Show me the technical findings 
before moving to content analysis.

Use Case 7: Podcast to TikTok Clip

Turning a long podcast into a short social clip normally means opening the video, downloading, finding the segment, trimming, converting to vertical, captioning, and exporting. Multiple tools. Multiple hours.

One prompt. Computer downloads the podcast, finds the segment by topic, extracts the clip, converts to vertical, captions it, and renders the final output.

Download the latest [PODCAST NAME] episode from [PODCAST URL]. 
Find the segment where they discuss [TOPIC]. Extract a 60-90 
second clip, convert it into vertical format for TikTok, and add 
captions with white text on a black background bar. Do not upload 
or publish anywhere without my approval. Show me the identified 
segment timestamp before extracting. Show me a preview before 
the final render.

View the Task

Use Case 8: Build Internal Tools in Plain English

Need a client intake form, a project tracker, or a simple CRM? Describe it. Computer builds it, styles it, and gives you a live link.

For tech users: Computer uses GPT-5.3-Codex to build full-stack apps with React frontends and Python backends, and pushes code directly to GitHub.

Build me a client intake form that collects a name, email, project 
type (dropdown with [TYPE 1], [TYPE 2], [TYPE 3]), budget range, 
and timeline. Save everything to a spreadsheet and send me a Slack 
notification every time someone submits. Provide a live shareable 
link. Single HTML file with inline CSS and JS. Clean, minimal 
design. Mobile responsive. Propose your approach before coding and 
wait for my confirmation. Show me the working form before you 
connect it to Sheets or Slack. Do not deploy without my approval.

How to Prompt Without Wasting Credits

I learned this the expensive way. Here’s the system I built so you don’t repeat my mistakes.

Computer is not one AI model. It’s an orchestration layer on top of 19 models. When you write a prompt, a routing system reads it, breaks it into subtasks, creates a separate sub-agent for each one, picks which model each agent runs on, and executes all of them.

Vague prompts cost real money. “Research the AI agent market and put together a report” might spawn four sub-agents. One scanning news. One pulling funding data. One analyzing strategies. One formatting. Each agent picks its own model. None of them asked whether you needed all four workstreams.

So how do you control a system that decides on its own how many agents to spin up?

I use a mental checklist before writing any Computer prompt. Five questions. I call it AGENT, but it’s a planning tool for YOU, not a formatting standard for Computer.

Computer reads natural language. It doesn’t need labeled sections. What it needs is clarity and constraints.

Before you write any prompt, ask yourself these five questions.

A: Assignment. What’s the one deliverable? Write the end goal of the task in a sentence.

G: Guardrails. What should Computer NOT do? This is the highest-leverage question. Every topic you don’t exclude is a rabbit hole. Every output type you don’t rule out (images, video, code) is a sub-agent. Every source you don’t cap is more credits burned.

Three things to always constrain. Scope: what to leave out and how far back in time. Agents: what types of output to suppress (no images, no video, no code if you just need text). Sources: cap the number. Ten well-chosen sources produce better results than forty scraped indiscriminately.

E: Engine. (Optional, experimental.) Can you specify which models to use? Computer lets you assign models to sub-tasks. I’ll be upfront: I haven’t confirmed this reliably overrides Computer’s default routing. The orchestrator makes its own decisions. But I mention it in some prompts because my results were more predictable with model hints than without. Try it. Check the model attribution in your results. See if it actually changes anything for your tasks.

N: Notify. When should Computer stop and check with you? This is the most important part after the deliverable itself. Always include: “Before you start, propose a plan and wait for my confirmation.” That one line forces Computer to show you what it’s about to do before it spends a single credit. You catch over-engineering before it costs anything.

T: Target. What does the output look like? Format, length, audience, delivery. Always set a length cap. Without one, synthesis agents keep writing until they exhaust the material. Always specify delivery. Computer connects to 400+ integrations and might try to send your output somewhere you didn’t intend.

Then write your prompt as a natural paragraph. Include all five elements. No labels needed. Computer reads English.

Here are three before-and-afters showing this in practice.

Before:

Research the AI agent market and create a comprehensive report 
on the major players, their pricing, features, and market position.

This prompt gave Computer no constraints. It created 6+ sub-agents, pulled sources on AI agents broadly (not just the market segment I wanted), generated comparison charts I never asked for, and spent credits on a “future predictions” section that added zero value.

After:

Build a comparison of the top 5 AI agent platforms (Perplexity 
Computer, Claude Cowork, OpenClaw, Manus, Operator) covering 
pricing, setup effort, and research quality. March 2026 data only. 
Exclude developer-only frameworks like LangChain, CrewAI, AutoGen. 
Maximum 8 sources per platform. No images, video, or code generation. 
Deliver as a markdown document with a comparison table at the top 
and a 300-word section per platform. 2,000 words max. Before you 
start, propose a task plan and wait for my confirmation.

Computer proposed a clean 3-phase plan. I removed one unnecessary research phase before it started. The output was tighter, more focused, and significantly cheaper.

Before:

Set up a system that monitors AI news and sends me a daily 
summary of what matters.

Computer created a monitoring agent, a summarization agent, a formatting agent, and a delivery agent. It pulled from dozens of sources per run and sent me everything, including noise I would have filtered out myself.

After:

Scan TechCrunch AI, The Information, Semafor Tech, VentureBeat AI, 
and Ars Technica AI for product launches and funding rounds from the 
past 24 hours. Only include product launches and funding rounds. 
Exclude opinion pieces, analysis, tutorials, and policy news. 
Maximum 5 items. Give me a numbered list with one sentence per item 
and a source link. Under 300 words. Only notify me if at least 2 
qualifying items are found. If fewer than 2, stay quiet.

The daily output went from a 1,500-word dump to a focused 200-word briefing. More useful. Far cheaper.

Before:

Build me a dashboard that shows my team's project status.

Computer interpreted “dashboard” as a full-stack application. It scaffolded a React frontend, wrote API routes, attempted to set up a database, and tried to deploy to Vercel. I wanted a single static HTML page.

After:

Build a single-file HTML project status dashboard showing 6 projects 
with name, owner, status (on track / at risk / blocked), and last 
updated date. Single HTML file only with inline CSS and JS. No 
backend, no API calls, no database, no deployment. No images or 
video. Dark mode, responsive. Exactly 6 project cards with those 
4 fields. Nothing else. Propose your approach before coding and 
wait for my confirmation. Do not deploy anywhere.

Computer built exactly what I asked for in one pass. No unnecessary infrastructure.

The pattern across all three. Same task. The specific version tells Computer exactly what to build, what NOT to build, and when to check with you before spending.

Three Mistakes That Drain Credits

Using Computer for tasks a cheaper tier handles. Computer spins up agents, models, and a full sandbox for every task. If Pro Search can answer it in 10 seconds, don’t waste Computer credits.

Not watching for retry loops. When Computer hits a failed connector, it retries. Then retries again. No hard stop. A broken step can silently drain hundreds of credits. Stay present for first runs. Automate after the workflow is proven.

Not calibrating before you scale. Run a lightweight version first. 3 sources instead of 10. Same structure, smaller scope. That tells you the full cost before you commit.

Quick Credit-Saving Tips

These are habits that save credits every day. Small things that add up fast.

Add a check-in step in plain English. Append this to any prompt: “Look, before you build anything, just double check with me what that looks like.” That one sentence prevents Computer from building things you didn’t actually want.

Be specific about what you want. “Research AI agents” is expensive. “Compare pricing of these 5 specific AI agents from March 2026” is cheap. Specificity reduces the number of sub-agents the orchestrator spawns.

Pay attention to system warnings. Computer has a built-in safeguard. When you give it a massive objective, the system pauses, warns you about credit consumption, and asks for approval. Don’t just click yes. Read the warning. Evaluate whether the scope is what you intended.

Keep your Custom Instructions short. You get 1,500 characters. Don’t use all of them. Unnecessary text increases your context window, which burns credits faster on every task.

Watch multi-chain tasks closely. Multi-chain prompting, app building, and long research sessions eat credits much faster than you’d expect. Monitor at perplexity.ai/account/usage.

3. Honest Limitations and Structural Risk

I need to be straight with you about some things nobody else is discussing.

It’s not a search engine. Treat it like one and you’ll burn credits on results Pro Search delivers in 10 seconds.

Connector reliability is uneven. Some work flawlessly. Others are flaky. OAuth tokens expire. The marketing says 400+. The number of battle-tested ones is smaller. Test before you automate.

Credit transparency is a real problem. No per-task pricing table. You find out the cost after. For a consumption-based product, that’s a significant gap.

And then there’s the structural risk. I wrote a full thread about this on X. Computer’s core reasoning runs on Claude Opus 4.6, built by Anthropic. Deep research runs on Gemini, built by Google. Lightweight tasks go to Grok, built by xAI. Long-context recall uses ChatGPT 5.3, built by OpenAI.

Perplexity built none of them. The product is a routing layer.

Every model provider is already building orchestration in-house. Anthropic ships Claude Code and Cowork. OpenAI has Operator. Google has Gemini with native tool use. The moment these models get good enough at everything, the routing layer becomes a feature someone else bundles for free.

Don’t sleep on it. It’s like OpenClaw for non-technical folks. And it’s buoying Perplexity’s growth. They’re back on Ramp’s list of fastest-growing software vendors for businesses.

[Bonus] Downloadable Takeaway

That’s it for today’s deep dive.

I wrote an extended version for paid members with seven more Computer workflows for PMs. Read it on Product Growth →

I watched Mike Krieger’s (former CPO at Anthropic, current leader of Labs) conversation with Dan Shipper and it was fascinating. Sharing my takeaways:

1/ The hard part of building flipped overnight. AI can go from zero to feature-complete in hours. That’s not the bottleneck anymore. The bottleneck is knowing what to cut. Claude will happily build you a monstrosity of overbuilt functionality if you let it.

2/ Stop building indoor trees. Krieger’s metaphor is perfect. Vibe coding lets you build an entire product indoors without ever exposing it to the wind of real users. If you’re 6 weeks into a build and nobody outside your team has touched it, you’re growing an indoor tree.

3/ The next generation of software is agent-native. Agent-native means the model can navigate and modify the product’s own primitives. Claude Code already knows about its own skills and can modify its own harness.

4/ Small teams are about to eat large teams alive. The unit of value is shifting from coordinated teams to autonomous individuals. Designers writing as much code as engineers. Scaling headcount too fast is now a net negative because models change so quickly that half your codebase might need to be deleted every few months.

The winning role is the GM who spikes on product sense and uses AI to manage the entire stack.

Watch Full Conversation

That’s all for today. See you next week,

Aakash

Share

P.S. Want my AI tool stack? Join my bundle. Want my job searching coaching? Apply to my cohort.

Everyone covered the ML side. But the pattern underneath works on anything you can measure. Skills, prompts, agents, email, page speed. I’ve spent the last two weeks pulling it apart, and today’s deep dive is your complete guide to using it.

But first, I cover Google Maps getting its biggest AI upgrade in over a decade, plus all the news that matters from the week.

Here’s what you need to know:

Ariso: Your AI Chief of Staff

I hired a chief of staff this year. It costs $0/month and lives in my Slack.

Ariso’s AI connects to my email, calendar, and docs, then handles meeting prep, follow-ups, and the busywork I used to lose hours to. It remembers every conversation and gets smarter the more I use it.

Built by cybersecurity founders, so your data is encrypted per-user and never trains their models.

I personally recommend this tool. Try it free at ariso.ai/aakash.

Try Ariso

Top News: Google Maps Just Got Its Biggest Upgrade in Over a Decade

OpenAI and Anthropic are building the best AI chatbots in the world. But Google just reminded everyone it owns the browser, the email, the maps, and the search engine.

Google announced two major features last week. Ask Maps is a Gemini-powered conversational layer that lets you type things like “My phone is dying, where can I charge it without having to wait in a long line for coffee?” and get real answers pulled from 300 million+ places and 500 million reviews. The second feature, Immersive Navigation, replaces the flat map with a full 3D view of buildings, terrain, stop signs, and lane markers. It analyzes Street View imagery and aerial photos using Gemini models to give you a real picture of what’s actually outside your windshield.

Here’s why this matters for builders. Google isn’t just shipping AI features in a chatbot. It’s embedding Gemini into the products 2 billion people already use every day. Maps, Gmail, Chrome, Android, Search. That’s not a model play. That’s a distribution play. And right now, no one else has that combination.

OpenAI can answer questions about the world. Google can answer questions about you.

The Other News That Mattered

• Travis Kalanick spent eight years building a ghost kitchen company. That was the cover story. He just renamed it Atoms and revealed he’s been building a robotics company across food, mining, and autonomous transport. Thousands of employees, 110 cities, nobody was allowed to put the company name on LinkedIn. The same board that pushed him out of Uber is now reportedly backing him.

• A tech entrepreneur in Sydney used ChatGPT to help design a custom mRNA cancer vaccine for his dying dog. Paid $3,000 to sequence the tumor DNA, ran mutations through AlphaFold, produced a personalized vaccine in under two months. The tumor shrank 75% while Moderna is spending billions to scale the same workflow for humans.

• Anthropic shipped Dispatch, a new Cowork feature that lets you control your desktop AI agent from your phone. Pair your phone with Claude Desktop, message from anywhere, come back to finished work. I’ve been running Cowork since launch and this was the missing piece.

• An engineer in Morocco open-sourced a phased array radar you can build from GitHub files. Commercial systems start at $250,000. The AERIS-10 does electronic beam steering and multi-target tracking at 20km range for roughly $5,000-$15,000 in parts. Open source collapsing defense contractor paywalls in real time.

Resources

• Claude now builds interactive charts, diagrams, and visualizations right inside your conversation. Not in a side panel. Not as a downloadable artifact. Ask it to explain compound interest and you get an interactive curve you can play with. This is the kind of feature that makes you stop opening Excalidraw for quick explanations.

• If you’re building Claude skills, I urge you to make them self-improving. Not just agents with skills, but agents with skills that get better over time. SKILL.md is clearly here to stay. The question now is whether your skills are static instructions or living systems that learn from their own outputs.

• I urge you to start using subagents. Codex just dropped subagent support. Instead of running everything in one thread, you can now spin up parallel agents that handle different parts of a task independently. If you’re not breaking complex tasks into coordinated subtasks yet, you’re leaving speed on the table.

Tools

• Google Stitch just got a massive update aimed directly at Figma’s throat. Google acquired Galileo AI, killed the $39/month subscription, and now gives you 350 design generations per month for free. The DESIGN.md export captures your entire design system as agent-readable markdown that flows into Cursor, Claude Code, and Firebase. Stitch is Maps for the design workflow.

• Claude Code now has Remote Control. Kick off a task in your terminal, pick it up from your phone while you take a walk. Claude keeps running on your machine and you control the session from the Claude app or claude.ai/code.

Fundraising

• Cursor is raising at a $50 billion valuation. In the same week, Elon pulled two of its product leaders into xAI to build the coding product missing from Grok’s stack. AI coding is now a $5B+ market and every frontier lab has a product printing money. xAI was the only one that didn’t.

The Ultimate Autoresearch Guide

Subscribe now

You’ve been settling. Your ad copy converts at 2%. Your onboarding emails get opened but nobody activates. Your AI-generated content needs three rewrites before it’s usable. You know all of this could be better, but testing 50 variations by hand isn’t a real option when you’re also trying to run a business.

Andrej Karpathy built the fix. He called it autoresearch, and 42,000 people starred it on GitHub in its first week. Fortune named it “The Karpathy Loop.”

Every article about it focused on the machine learning angle. They all missed the bigger story. The pattern underneath works on ad copy, email sequences, video scripts, job posts, and anything else where you can define what “good” looks like. I spent two weeks pulling it apart, going through Karpathy’s repo, the community forks, and the real-world applications people are building on top of it. This guide shows you how to use it on the things you actually make every day.

You don’t need a GPU. You don’t need to know anything about ML.

The tech world moved on after the ML headlines. That was premature. The core loop has nothing to do with training neural networks. It takes whatever you’re producing with AI, runs 100 variations while you sleep, keeps the ones that score better, throws away the ones that don’t, and saves the improved version for when you wake up.

1. What autoresearch actually does

Karpathy builds small language models for fun. The workflow looks the same whether you’re training a model or tuning a prompt: make a change, test it, check if the output got better, keep or discard, repeat. On a good day he’d get through 8-10 cycles. Most of that time was spent waiting.

He automated the entire thing.

Three files power the system. One file the agent edits (the thing being optimized). One file the agent can never touch (the scoring criteria, because if it could, it would just make the test easier). And one instruction file written by the human that tells the agent what to try and how to behave.

Each cycle takes about 5 minutes. The agent reads the current version, decides what to change, makes the edit, runs the test, and checks a single number. Score went up? The change gets committed to git. Score went down or stayed flat? git reset wipes it clean. Next round starts.

Twelve cycles per hour. A hundred overnight. About $25 in compute.

Karpathy ran it for two days on code he’d already optimized by hand for months. The agent found 20 improvements he’d missed, including a bug in his attention implementation. All 20 stacked and transferred to a larger model. 11% faster.

Tobi Lutke, the Shopify CEO, tried it the same night. 37 experiments while he slept. His smaller model started outperforming his larger one. Then he turned the loop on Liquid, Shopify’s templating engine, and got 53% faster rendering from 93 automated commits.

The ML part was the proof of concept. The loop itself is the product.

2. Why this works beyond ML

Forget the neural network part. The loop runs on any output where three conditions hold.

You can score the output with a number. Not a feeling, not a vibe check. Binary questions work best. “Does the ad headline mention a specific pain point?” Yes or no. Run 30 outputs through 5 binary questions and you get a percentage. That percentage is your metric.

The scoring runs without a human. Your coding agent builds an evaluation script that generates outputs, grades them against your questions, and prints the score. No one needs to sit and read 30 ad variants at 2am. The machine scores them and moves on.

Only one file changes per round. Your prompt, your template, your skill file. The agent edits this single file. Everything else is locked. This constraint is what makes the loop work. If the agent could edit the scoring criteria too, it would just make the test easier instead of making the output better.

All three true? The loop works. Any one missing? It won’t. That’s the whole filter.

Your job is choosing what “better” means and writing the yes/no questions that define it. The agent handles the 50 rounds of iteration you were never going to do manually. I run this in Claude Code, but Cursor, Windsurf, Codex, and Antigravity all work. Any coding agent that can read files, edit files, and use git.

3. How to set it up

Three steps. Five minutes of setup, then the agent takes over.

Step 1: Clone the repo

git clone https://github.com/karpathy/autoresearch ./autoresearch-reference
git init && git add . && git commit -m "setup"

This gives your coding agent a reference copy of Karpathy’s convention. If you already have a prompt or skill file you want to improve, drop it in the same folder.

Step 2: Write your eval criteria

This is the human’s job and the part where most people succeed or fail.

You’re writing yes/no questions that describe what a good output looks like. Here’s what I used for a landing page copy skill:

1. Does the headline include a specific number or measurable result?

2. Is the copy completely free of buzzwords?

3. Does the CTA use a specific action verb tied to the product outcome?

4. Does the first sentence after the headline name a specific pain point?

5. Is the total copy between 80 and 150 words?

Five binary questions. The agent generates 30 outputs per round, scores every one, and produces a single percentage.

Don’t know where to start? Just describe what goes wrong when your current output fails. “The headlines are generic, there’s too much jargon, the CTA is weak.” Claude Code will convert those complaints into binary scoring criteria.

Three to six questions is the range. Below three and the agent finds loopholes. Above six and it starts gaming the checklist instead of improving what the output actually sounds like.

Step 3: Tell Claude Code to run the loop

Read autoresearch-reference/program.md to understand Karpathy's 
autoresearch convention. I want to optimize [my skill / prompt / 
system prompt] using these scoring criteria: [your yes/no questions]. 
Test against these inputs: [2-3 realistic scenarios]. 

Set up the evaluation harness, run the baseline, then enter the 
autoresearch loop. One change per round. Keep if score improves, 
revert if it doesn't. Never modify the evaluation after creating it.

One prompt. Claude Code reads the convention, builds the eval script, generates a baseline score, and starts iterating.

4. Six use cases you can run tonight

Use case 1 - Self-improving skills

You need to steal this autoresearch skill optimization use case immediately.

I ran this on a landing page copy skill with the 5 eval criteria. The baseline came back at 41.3%

Round 1: Claude Code saw that “specific number in headline” was failing 80% of the time. It added a rule to the skill requiring specific numbers. Score jumped to 68%. Kept.

Round 2: Buzzwords were the next biggest failure. It added a banned words list. 79%. Kept

Round 3: CTAs were still generic. It added a worked example showing what a strong CTA looks like. 90%.

Round 4: It tried tightening the word count. Score dropped to 82% because shorter copy hurt the CTA.

41% to 92%. 4 rounds. Three changes kept, one reverted. The eval caught the regression automatically.

Start with the skill that frustrates you most. The one where you re-run 3 times before getting usable output.

Use case 2 - Cold email that optimizes itself

Your outreach reply rate is at 2-3%. You’ve tried manual A/B tests but it takes weeks to get signal.

Eval criteria: Under 75 words? References the prospect’s specific role? Ends with a concrete question? First two sentences include a specific number or result?

The agent discovers what human copywriters learn over years: shorter beats longer, specificity beats vague, risk-reversal beats feature lists. Eric Siu, founder of ad agency Single Grain, is already building this for his clients. His framework replaces the training script with the email itself. The agent modifies the subject line or CTA, measures reply rate, keeps or discards. Siu’s math: most marketing teams run ~30 experiments a year. This loop runs 36,500+.

MindStudio documented teams using this pattern seeing reply rates move from a 2-4% baseline to 8-12% within four to six weeks. For the live version, connect to your email platform’s API (Instantly, Apollo, Lemlist) and deploy to GitHub Actions on a cron schedule.

Use case 3 - Ad copy that tests itself

You’re running Meta or Google ads and testing 2-3 headline variants per week by hand. That’s maybe 150 tests per year. Autoresearch runs 100 overnight.

Eval criteria: Is the headline under 40 characters? Does it name a specific pain point in the first 5 words? Does the body mention a number or timeframe? Is the CTA a verb phrase, not a noun phrase (”Start saving” not “Savings plan”)? Does it avoid every word on your brand’s banned list?

Siu’s framework at Single Grain applies here directly. The agent iterates on your ad copy template the same way it iterates on a training script. Each round generates 30 ad variants, scores them all, keeps the template change if the average score went up, reverts if it didn’t. MindStudio’s guide walks through the full implementation for landing pages and ad creatives, including how to connect the loop to your CMS via API. You wake up to a template that produces compliant, specific, action-oriented copy on the first try.

Use case 4 - Short-form video scripts

Your Reels and TikToks hit when the hook lands and miss when it doesn’t. You know this intuitively but you test hooks by posting and waiting 48 hours for the algorithm to tell you.

Eval criteria: Does the opening line create a specific curiosity gap? Is there a surprising claim in the first 2 sentences? Does it follow a single narrative thread (not 3 loosely connected points)? Is the script under 60 seconds when read aloud? Does the last line set up the next piece of content?

The agent can’t predict virality. But it can enforce the structural patterns that correlate with watch time. witcheer, Head of Growth at crypto platform Yari Finance, ran autoresearch overnight on a Mac Mini M4 and found that “the model got better by getting simpler.” The same principle applies to scripts: the agent strips filler, tightens hooks, and enforces a single thread. The creative ideas are still yours. The structural consistency is the machine’s job.

Use case 5 - Onboarding email sequences

You wrote your 5-email onboarding sequence six months ago. Open rates are fine. But activation is stuck at 30% and you’ve been too busy to rewrite them.

Eval criteria: Does each email have exactly one CTA? Is the email under 100 words? Does the subject line include a specific benefit or number? Does the first sentence reference the user’s last action (not a generic greeting)? Does it avoid “we” in the first paragraph (focus on them, not you)?

One file changes: your email template. The eval generates 30 onboarding sequences per round, scores every email in every sequence, and gives you an aggregate. If you want to go further, 35 agents on the Hyperspace network ran 333 experiments in a single night, completely unsupervised. Agents on cheap hardware found optimizations the expensive setups missed, because constraints forced creative solutions. The same distributed approach works on email: spin up multiple agents testing different parts of the sequence in parallel.

Use case 6 - Job descriptions

Most job posts are written once and never optimized. They’re too long, full of jargon, and read like a legal document. Then you wonder why qualified people don’t apply.

Eval criteria: Is it under 400 words? Does it name a specific project the hire will work on in the first 90 days? Is it free of internal jargon and acronyms? Does it include compensation range? Does the first sentence describe impact, not responsibilities?

This is one of the cleanest autoresearch applications because the eval criteria are obvious and binary. The agent strips jargon, forces specificity, and front-loads impact. After a few rounds you have a template that produces job posts people actually want to read. Forward this to your recruiting team and watch application quality change.

5. What most people get wrong

After spending two weeks with this, I’ve seen the same three mistakes kill autoresearch runs before they start.

Mistake 1: Vague eval criteria.

“Is the copy compelling?” is not a question an agent can score. “Does the headline contain a specific number?” is. Every criterion must be binary. Pass or fail, yes or no, no sliding scales. The moment you introduce a 1-7 rating, the agent starts optimizing for 4s that technically score well but read like garbage. If you can’t explain to a stranger how to score it in one sentence, rewrite the criterion.

Mistake 2: Touching the eval mid-loop.

Your baseline comes back at 35% and your first instinct is to soften the criteria so the number looks better. This is the single fastest way to waste a run. The eval is locked the moment the first round starts. If you change the scoring mid-loop, every previous round’s data becomes meaningless. You’re comparing apples to oranges across your own experiment log. If the eval is wrong, stop the run, fix it, and restart from scratch. Karpathy made prepare.py read-only for exactly this reason.

Mistake 3: Too many criteria.

Seven or eight eval questions feels thorough. In practice, the agent starts gaming the checklist. It finds combinations of changes that technically pass 7 out of 8 criteria but produce outputs that don’t read like something a human would write. Three to six criteria is the sweet spot. Enough to prevent shortcuts, few enough that the agent optimizes for real quality instead of checklist compliance. When in doubt, start with 3 and add one at a time if the output has a failure mode you didn’t cover.

The pattern is simple. The eval is the hard part. Get that right and the loop does the rest. I explained the broader vision here.

Pick one thing you’ve been meaning to improve. Write 3-5 yes/no questions that define what “good” looks like. Let the loop run tonight. Wake up to something that works.

That’s it for today’s deep dive.

I wrote an extended version for PMs with six more use cases (page speed, agent system prompts, production agents, content feedback loops, and more) plus the experiment log strategy that makes this compound over time. Read it on Product Growth →

I watched David Senra’s conversation with Marc Andreessen and it’s one of the best founder interviews I’ve heard this year. Sharing my takeaways because this one hit different.

1/ Great founders have near-zero introspection. Andreessen says the most successful entrepreneurs spend almost no time self-analyzing or dwelling on the past. Sam Walton had zero inner dialogue about his feelings. He just woke up every day and kept building Walmarts. The whole concept of sitting around analyzing yourself is newer than we think. Before Freud, nobody did it. They just went out and built things.

2/ Founders beat professional managers every time during rapid change. This is a core a16z thesis. Formally trained managers can maintain status-quo operations fine. But when industries undergo massive technological shifts, they fail. It is much more effective to start with an innovative founder and teach them how to manage than to expect a professional manager to invent like a founder. If you’re a PM watching AI reshape your industry, this should feel personal.

3/ Every major technology faces a moral panic. Every single one. The bicycle got doctors warning about “bicycle face.” The printing press, jazz, TV, video games, social media. Same pattern. The world is far more malleable than people realize, and founders who push past the panic with maximum effort can force the world to recalibrate around their ideas. AI is getting the same treatment right now.

4/ Elon Musk runs a management style nobody else is willing to try. Traditional CEOs sit behind layers of management that filter the truth. Andreessen compares it to IBM’s historical “big gray cloud.” Musk does the opposite. Every week he maps the exact production bottleneck across each of his companies, bypasses every management layer, and works hands-on with frontline engineers to fix problems directly. That speed is the competitive advantage, not the rockets or the cars.

Watch Full Conversation

That’s all for today. See you next week,

Aakash

Share

Karpathy just open-sourced a research lab that runs while you sleep. Netflix dropped $600M on a 16-person startup. And human brain cells learned to play Doom.

In today’s issue, I cover all the news that matters from the week.

And for today’s deep dive, I go deep on the SOTA Image Gen model that dropped 2 weeks ago, Nano Banana 2:

Nimbalyst: The Visual Workspace for Claude Code and Codex

Today, you describe what you want in a terminal, the agent writes files to disk, and you open them somewhere else to see what happened. Like giving renovation instructions through text messages.

Now imagine working in the same visual workspace as your agent. Planning, coding, diagramming, tracking tasks, all connected. You see what the agent is doing and approve it. The agent sees what you’re doing and uses it.

This is Nimbalyst. Download it for free here:

Try Nimbalyst

There’s a billion AI news articles every week. Here’s what actually mattered.

The Week’s Top News: Karpathy Just Turned One GPU Into a Research Lab

Andrej Karpathy dropped autoresearch last weekend, and the internet went sideways. The repo crossed 27,000 stars within days.

Here’s what it actually is: A 630-line Python script that puts an AI agent in charge of running ML experiments on a single GPU:

• The human writes a Markdown file with high-level instructions.

• The agent modifies the training code, runs a 5-minute experiment, checks if the result improved, commits the good ones, and loops back. Twelve experiments per hour. Around 100 while you sleep. No breaks, no forgetting, no lab politics.

This matters because it changes who can do ML research. A PhD student at a mid-tier university with one GPU can now run 100 experiments overnight. A startup that can’t afford a research team gets automated ablation studies across hyperparameters, architectures, and training schedules while the founders sleep. A solo developer testing a new loss function gets 12 runs per hour instead of doing one, checking Slack, doing another, losing the thread.

The use cases go beyond academic papers. Product teams can A/B test model variants against real evaluation criteria. Infrastructure teams can automate regression testing across model versions. Fine-tuning shops can let the agent grind through LoRA configs until it finds the one that actually moves the benchmark.

Right now it’s one agent on one GPU. Karpathy already outlined what comes next: hundreds of agents running parallel research branches, reading each other’s results, combining findings.

A 630-line script just turned every GPU on the planet into a potential research lab.

The Other News That Mattered

• Netflix acquired Ben Affleck’s AI startup InterPositive for a reported $600M, one week after walking away from an $83B Warner Bros deal. InterPositive trains models on a production’s own dailies and lets filmmakers fix continuity errors, relight shots, and remove wires in post. Netflix spends $20B on content annually. Even a 10% cut in post costs pays for this many times over.

• Cortical Labs grew 200,000 human neurons on a chip and taught them to play Doom. An independent developer with zero biotech experience built the integration in a week using a Python API. Pong took 18 months. Doom took a week. Your brain runs on 20 watts. A GPU cluster draws megawatts. The energy economics here are not close.

• Claude Code launched local scheduled tasks via a new /loop command. You describe a task, set a cadence, and Claude runs it in the background while your computer stays on. Check error logs every few hours, create PRs for fixable bugs, generate morning Slack summaries.

• ChatGPT now has 140 million people using it weekly for math and science, and they just added interactive visual modules for 70+ topics. Adjust a variable in Ohm’s law, watch the graph update in real time. Khan Academy spent 18 years reaching 2 million AI tutor users. OpenAI has 70x that and just started adding the interactive layer.

Resources

• Within a year, every company over 50 people will have at least one full-time Agent Builder. This piece on the rise of the Agent Builder is the clearest breakdown I’ve seen of what that role actually looks like.

• If your lawyer isn’t using AI like this, find a new lawyer. Zack Shapiro’s piece on running a Claude-native law firm is one of the best real-world examples of AI replacing entire workflows, not just speeding them up.

• OpenAI’s newer models hallucinate 3x more than the ones they replaced. o1 hallucinated 16% of the time. o3 hit 33%. o4-mini hit 48%. They published a paper explaining exactly why they can’t stop it and the answer is uncomfortable: the system that produces intelligence and the system that produces hallucinations are the same system.

Tools

• Google’s new Workspace CLI is the fix nobody knew they needed. One developer found his Google MCP setup loading 37,000 tokens into context before the agent even started working. The CLI sidesteps that entirely. Same capabilities, fraction of the overhead. 100+ agent skills. As agents get smarter, the bottleneck shifts from “can it access the tool” to “how much context does it cost.” CLIs win that math every time.

Fundraising

• Fei-Fei Li raised $1B for World Labs. Now Yann LeCun raised $1.03B for AMI Labs, a company that is 4 months old with zero product and zero revenue. Both closed within three weeks with overlapping investors. The bet they’re both making: large language models have hit a ceiling, and the future belongs to world models that understand actual physics. If they’re right, a lot of companies building physical-world applications on top of current LLMs picked the wrong foundation.

The Ultimate Guide to Nano Banana 2 🍌

Nano Banana 2 dropped two weeks ago and I immediately started testing.

What I found genuinely surprised me. This isn’t a cheaper version of Pro (the last drop in November). It’s a better model:

• PDF input up to 50MB

• Google Image Search grounding

• Real-time web search grounding

• A context window double the size of Pro

• A 512px resolution tier that makes high-volume economics work in a way they never did before

• And a thinking mode that reasons through compositions before generating - which changes first-pass quality more than I expected

I ran it through two weeks of real work: competitive analysis, prototype iteration, deck prep, stakeholder docs, UGC seeding.

Here’s what I learned, and every prompt worth stealing.

Subscribe now

How to Access Nano Banana 2

30 seconds to set up.

1. Go to gemini.google.com

2. Click on “Tools” and then “🍌 Create image”

3. Switch to the Thinking model

   

   That’s it. You’re on Nano Banana 2.

The 13 Most Useful Workflows

I tried out 100+ things and found the 13 you should care about:

1/ The Frankenstein Mockup

You’re walking into a product review and you need to show what a new feature could look like. Design is tied up for days even if you are using AI. You have your app’s nav in one tab, a competitor layout you like in another, your brand color palette somewhere, and a whiteboard sketch from last Tuesday.

Attach all four. That’s the entire prompt setup.

Using the 4 attached reference images — [1] my app's current navigation bar, 
[2] a competitor's card layout I want to adapt, [3] our brand color palette 
screenshot, [4] my whiteboard sketch of the new feature — generate a high-fidelity 
UI mockup of [FEATURE NAME]. 

Match the component style from image 1, adapt the card structure from image 2, 
apply the exact colors from image 3, and reflect the rough layout from image 4.

Render as a mobile screen, 9:16 aspect ratio, white background, clean and 
implementation-feasible. Use realistic UI elements, no placeholder icons.

2/ Research Insight to Shareable Infographic

You’ve done the user research. Nobody reads the 40-page report. You need something you can drop in Slack, paste into a deck, or print for a workshop before the next standup.

Nano Banana 2’s text rendering is the reason this works cleanly now. Earlier versions of the model couldn’t hold sharp legible type across the whole image. This one does.

Generate a clean, professional infographic summarising the following 
research insight:

[PASTE YOUR KEY FINDING HERE — 2-3 sentences with the core stat]

Layout:
- Top: Bold headline stating the core finding
- Middle: 3-4 data points in labelled visual blocks, each with a large 
  typographic number as the anchor
- Bottom: One "what this means for us" callout in a distinct colour band

White background, [BRAND COLOR] as primary accent, clean sans-serif 
typography, no stock photo elements, data visualisation only. 
Aspect ratio 4:5, 2K resolution.

3/ The Competitive Visual Grid

You need to show leadership how three competitors are handling the same problem. Building a real comparison takes a designer half a day and your deck is in two hours.

Describe what you need. Let Nano Banana 2 search for the current UI and render the comparison.

Search the web for the current [SPECIFIC SCREEN — e.g., "onboarding empty 
state" / "checkout flow" / "settings page"] of [COMPETITOR 1], [COMPETITOR 2], 
and [COMPETITOR 3].

Then generate a clean side-by-side comparison grid showing how each 
product handles [SPECIFIC UX PATTERN].

Format as a 3-column grid with each competitor's name as a bold column 
header. Under each, show a simplified but realistic mockup of their approach 
with 2-3 annotation callouts highlighting the key design decisions.

White background, 16:9 aspect ratio, minimal styling, annotation callouts 
in small grey labels with arrows. This is for an internal strategy presentation.

4/ PRD to Visual

You’ve written a solid spec. Nobody reads it end-to-end. You spend the first 20 minutes of every product review re-explaining the concept before the actual discussion starts.

Feed the PDF. Let it generate the visual summary instead of you rebuilding the explanation every time.

I'm attaching my product requirements document as a PDF. Read it fully, 
then generate a clean visual overview of [FEATURE NAME] that shows:

1. The user problem being solved (before/after comparison)
2. The proposed solution as a labelled product diagram
3. The key user journey in 3-4 annotated steps

White background, [BRAND COLOR] accents, bold sans-serif headers, 
clean connecting arrows between elements.

Design this to be readable at 1080px width without zooming — 
it'll be embedded in Notion or shared in Slack.

5/ The Surgical Fix

You generated a mockup that’s 80% right. One element is wrong. You don’t want to start over.

This is what semantic masking is for. Describe the region by what it contains, not by coordinates. Nano Banana 2 changes only that and leaves the rest exactly as it was.

Keep the entire image exactly as it is. Make only this one change:

[SPECIFIC EDIT — e.g., "Replace the blue primary button in the bottom 
right with a button using hex color [YOUR BRAND HEX]. Change the button 
label to '[NEW LABEL]'. Keep the size, position, and shape identical."]

Do not change the background, layout, typography, icon placement, 
or any other element. Only modify what I described above.

6/ App Store Lifestyle Screenshots

App store screenshots that show the product being used by a real person in a real context outperform clean product shots. A phone on a white background tells someone what your app looks like. A person using your app at a coffee shop tells them what their life looks like with it.

Most teams don’t do lifestyle screenshots because they require a photoshoot. With Nano Banana 2, you skip the photoshoot entirely.

Generate a realistic lifestyle App Store screenshot showing 
[YOUR APP NAME] being used in an everyday context.

The person: [AGE, GENDER, DESCRIPTION — e.g., "a woman in her 
early 30s, calm, present, wearing casual home clothes"]

The setting: [ENVIRONMENT — e.g., "sitting cross-legged on a sofa 
by a large window, morning light, ceramic mug nearby"]

Camera angle: the person is looking DOWN at their phone naturally. 
The phone is NOT held up toward the camera — it sits at a natural 
angle in their hands or lap. Frame from eye level so the face is 
the clear primary subject. The phone screen should be readable 
but secondary to the human.

The screen shows: [DESCRIBE YOUR KEY UI — e.g., "a dark navy 
meditation screen with a soft glowing circle and 'Breathe in...' text"]

Shooting style: natural light, f/1.8 shallow depth of field, 
face in sharp focus, background soft. Editorial feel, not a product shoot.

Output at 9:16, 2K resolution.

7/ Feature Launch Social Asset

You just shipped something. You want a LinkedIn post that stops the scroll. You don’t want another screenshot with a gradient background. You want a person — your person, your user — in the moment your feature solves their problem.

Generate a realistic photo-style image for a LinkedIn feature launch post.

The scene: [DESCRIBE THE HUMAN MOMENT — e.g., "a product manager in 
her early 30s, sitting at a standing desk in a modern open office, 
looking at her laptop with a visible expression of relief — she just 
found the thing she was looking for"]

Demographic: [AGE RANGE, GENDER, ETHNICITY — e.g., "South Asian woman, 
late 20s to mid 30s, professional but not overly formal"]

Setting: [CONTEXT — e.g., "daytime, natural light from a nearby window, 
laptop open, coffee cup on the desk, slightly cluttered workspace — 
looks real, not staged"]

The laptop screen should show [YOUR PRODUCT SCREEN — attach a screenshot 
or describe the key UI moment]. It should be readable but not the focus 
— the human reaction is the focus.

Shooting style: candid documentary feel, slightly shallow depth of field, 
natural colour grading. Looks like a real photo a colleague took, 
not a commercial shoot.

Output at 4:5 (LinkedIn optimal), 2K resolution.

8/ Localized GTM Assets in One Prompt

You’re launching in India, the UAE, and Germany simultaneously. The old way means three separate briefings, a localization vendor, two weeks of back and forth, and assets that still don’t quite feel right because the translator and the designer never spoke to each other.

Nano Banana 2’s text rendering and multilingual support changes this. One prompt generates the same launch banner in Hindi, Arabic, and German — correct typography, right-to-left layout for Arabic, cultural context baked in. What used to be a two-week vendor process is a single generation.

Generate 3 versions of a launch banner for [PRODUCT/FEATURE NAME].

All 3 versions should be visually identical in layout and design, 
but rendered in different languages with culturally appropriate typography:

Version 1: Hindi — Devanagari script, warm and celebratory tone
Version 2: Arabic — right-to-left layout, Arabic script, 
           ensure the design mirrors correctly for RTL reading
Version 3: German — clean sans-serif, precise and professional tone

Banner layout:
- Top: Product logo placeholder
- Center: Hero headline "[YOUR LAUNCH MESSAGE]" in each language
- Below: Subheading "[YOUR VALUE PROP — 1 line]" in each language  
- Bottom: CTA button "[YOUR CTA TEXT]" in each language

Design: [PRIMARY COLOR] background, [ACCENT COLOR] for CTA, 
white text, [FONT STYLE] where Latin script is used.
Same visual hierarchy across all 3 versions.

Output each at 16:9, 2K resolution.

9/ Material and Texture for Hardware or Physical Product PMs

If you’re shipping something physical — a device, packaging, a consumer product — you know the pain of getting a photorealistic render for an investor deck or manufacturer brief without a 3D rendering team.

Nano Banana 2’s materiality framework is built for this. You’re not describing “a white box.” You’re describing “recycled kraft paper with an embossed logo, matte surface, slight texture visible under directional studio light.” It renders the difference.

Generate a photorealistic product render of [YOUR PRODUCT — e.g., 
"a wireless earbud charging case" / "a supplement bottle" / 
"a smart home device"].

Material: [EXACT MATERIALS — e.g., "matte anodized aluminum body, 
frosted glass top panel, brushed finish on the edges"]

Surface: [TEXTURE DETAIL — e.g., "subtle grain on the aluminum, 
fingerprint-resistant coating, logo embossed 0.5mm depth, centered"]

Lighting: Three-point softbox — main light upper left, soft fill 
right, rim light from behind to define edges. No harsh shadows.

Background: [e.g., "seamless warm white studio sweep" / 
"dark slate grey for premium feel"]

Composition: Product centered, angled 15 degrees, medium shot, 
full product visible. f/2.8 shallow depth of field.

Output at 1:1, 4K resolution, investor deck ready.

10/ Product Hunt Launch Hero Image

Your Product Hunt hero is the first thing 50,000 people see on launch day. Most founders put a screenshot or a gradient with their logo. The products that hit #1 almost always have a specifically composed, specifically lit hero that looks intentional.

You don’t need a designer for this anymore.

Generate a cinematic hero image for a Product Hunt launch.

Product: [NAME] — [ONE LINE DESCRIPTION]

Concept: [VISUAL IDEA — e.g., "close-up of hands typing on a laptop 
at night, screen glowing with the product dashboard, everything else 
in darkness — conveys focus and late-night momentum"]

Lighting: [STYLE — e.g., "single cold key light from the left, 
high contrast, screen glow as secondary light source — 
cinematic, not commercial"]

Camera: [LENS + ANGLE — e.g., "low angle looking up at the screen, 
wide-angle lens, slight distortion — makes the product feel significant"]

Color grade: [MOOD — e.g., "muted cool tones, slight blue shift, 
cinematic desaturation — modern and focused, not playful"]

No text in the image. Output at 16:9, 4K resolution.

11/ Product Mockup Shots Without a Photoshoot

You need a hero image for a press release, a launch tweet, or a pitch deck. No photoshoot budget. No time to brief a designer.

Describe the material. Specify the lighting rig. Define the surface. Nano Banana 2 renders a studio-quality product shot in 30 seconds that looks like it cost $3,000 to produce.

Generate a studio product shot of [YOUR PRODUCT — e.g., 
"iPhone 16 Pro showing the app's onboarding screen" / 
"MacBook Pro showing a SaaS dashboard" / 
"a physical device on a clean surface"].

Material: [e.g., "Space Black iPhone, titanium frame, matte back, 
screen showing the app's key moment"]

Lighting: Three-point softbox — [MOOD — e.g., "warm main light 
upper left, soft fill from right, subtle rim light from behind. 
Clean commercial product photography look."]

Background: [e.g., "seamless off-white studio sweep" / 
"dark textured concrete for premium feel" / 
"warm walnut wood for lifestyle feel"]

Composition: Product centered, angled 20 degrees, full product 
visible, no cropping. f/2.8, background slightly soft.

Output at 16:9, 4K resolution, press-ready.

12/ Personal Brand Headshot Style Transfer

Every founder and PM needs a headshot that doesn’t look like everyone else’s grey-gradient LinkedIn photo. The problem is getting a specific visual style — bold background, high contrast, editorial feel — applied to your actual face without booking a photographer who shoots in that style.

Nano Banana 2’s multimodal input changes this. Upload your photo. Upload a style reference. Tell it exactly what to transfer and what to leave alone.

The thing most people get wrong: they upload both images and just say “make mine look like this.” The model blends the two faces and the output looks like neither person. You need to explicitly separate what transfers from what stays.

Generate a styled headshot using the attached reference images.

Image 1 (attached): my photo — use my face, hair, and likeness exactly. 
Do NOT alter my facial features or expression.

Image 2 (attached): style reference — use the visual style only. 
Do NOT copy the face of the person in this image.

Transfer from the reference: [LIST WHAT TO TRANSFER — e.g., 
"background color and treatment, lighting style, contrast level, 
black-and-white vs color treatment, crop and framing style"]

Keep exactly: my face, my hair, my skin tone, my features, my expression.

Output at 1:1, 2K resolution.

13/ Research Insight to Styled Infographic

Your user research is dying in Google Docs. Nobody opens the synthesis deck. Nobody reads the Notion page.

The data is good. The format is the problem.

Nano Banana 2’s text rendering turns that same data into a shareable infographic in one prompt. But skip the clean corporate version the one that gets shared is the sketchnote doodle style. Hand-drawn, lined notebook paper, highlighter on the key phrases.

Here’s the prompt:

Generate a sketchnote doodle infographic in the style of a 
hand-drawn notebook page.

Background: lined notebook paper — horizontal ruled lines 
visible across the full image, warm cream paper color, 
spiral binding visible on the left edge. 
Looks like a real opened notebook.

Style: hand-drawn black ink doodle aesthetic throughout — 
imperfect lines, slightly uneven hand-lettered text, 
simple illustrative icons drawn per section. 
NOT clean vector. NOT corporate slides.

Layout: 2x3 grid of 6 equal sections with dashed border boxes. 
Banner headline at the top center in a hand-drawn ribbon shape. 
Small footer text at the bottom center.

Headline (bold hand-lettered caps inside ribbon banner):
"[YOUR HEADLINE — e.g., 6 THINGS EVERY PM SHOULD DO 
BEFORE WRITING A PRD]"

Content — 6 sections, one per box:
Section 1: [NUMBER + TITLE] — [1 LINE INSIGHT] — [DOODLE DESCRIPTION]
Section 2: [NUMBER + TITLE] — [1 LINE INSIGHT] — [DOODLE DESCRIPTION]
Section 3: [NUMBER + TITLE] — [1 LINE INSIGHT] — [DOODLE DESCRIPTION]
Section 4: [NUMBER + TITLE] — [1 LINE INSIGHT] — [DOODLE DESCRIPTION]
Section 5: [NUMBER + TITLE] — [1 LINE INSIGHT] — [DOODLE DESCRIPTION]
Section 6: [NUMBER + TITLE] — [1 LINE INSIGHT] — [DOODLE DESCRIPTION]

Highlighter accents: yellow (#FDE68A) and pink (#FBCFE8) — 
use on the 1-2 most important words per section only. 
Looks like someone went back over it with a real highlighter.

Footer in small neat handwriting at the bottom center:
"Subscribe to aibyaakash.com"

Output at 4:5, 2K resolution.

3 Things I Learned the Hard Way

Name what you want preserved, not just what you want changed. When I’m doing a surgical edit, “keep everything else identical” is not specific enough. Name the elements: background, layout, typography, colour of every other button. The more explicit you are about what stays, the more precise the edit.

Tell it where the output is going. “For a Notion page,” “for a 16:9 slide,” “for Slack” - this changes how Nano Banana 2 handles margins, text sizing, and resolution. I skip this when I’m rushing and then I’m always resizing manually afterward.

512px is your cost control lever. For high-volume generation, generate at 512px and upscale separately. You get the Nano Banana 2 reasoning quality at Nano Banana 1 pricing. The difference between 512px and 1K is invisible at screen resolution in most presentations.

That’s the full Nano Banana 2 playbook. Save the prompts, try the Frankenstein mockup first, and let me know what you end up building.

I watched Dylan Patel’s latest interview on the future of AI and chips. Four things stayed with me:

1. Agent orchestration is rewriting white-collar work. Claude Code and tools like it are no longer just coding assistants. They’re becoming full orchestration systems that let non-programmers build complex software and data pipelines in natural language. Companies using this are operating 10x faster than before. The junior developer market is getting nuked in the process.

2. The US is losing the military AI race. The US military is running on older classified models. China is deploying the newest ones for autonomous drone swarms right now. And while Anthropic is pushing back on government demands that could enable mass surveillance, OpenAI has already moved in to secure Department of War contracts.

3. Hardware constraints are locking in cloud AI dominance. Global shortages in memory and 3nm/5nm silicon are already forcing major mobile companies to cut smartphone production. The industry is routing everything scarce toward data center clusters. Local AI inference is becoming more expensive and less practical, not less. Closed-source cloud wins by default.

4. Scaling laws are still intact, and the ASI race is accelerating. Pre-training and RL scaling are both holding. Costs keep dropping year over year, capabilities keep jumping. Anthropic is the consensus leader right now based on revenue and model quality. OpenAI remains the contrarian pick to reach ASI first.

Watch Full Conversation

That’s all for today. See you next week,

Aakash

PS: You are receiving this email because you were subscribed to Aakash’s Product Growth Newsletter, from where the first 16 issues of the AI Update were sent. I have now given it its own home.

Share

Every single person I’ve turned on to Claude Code and Claude Cowork has been happy I did. I really think it’s not an exaggeration to say:

They’re the most important new professional software created in years.

But the issue everyone runs into is starting with them. Today, I’m open-sourcing everything you need to start with them well. So you have no excuse!

That’s today’s deep dive.

But first, there’s been an absolute torrent of interesting AI news. I’ve crossed 200M impressions on X alone the past month covering it. And today, I’m bringing you the web’s highest signal summary of the last week. But first…

Nimbalyst: The Visual Workspace for Claude Code and Codex

If you’re using Claude Code in the terminal, you’re missing half the picture.

Nimbalyst gives you visual editors for everything your agent touches. Markdown, CSV, mockups, diagrams, data models, code. You see exactly what the agent reads and edits, review changes as diffs, and approve before anything lands. No more switching between terminal, VS Code, and Figma:

Try Nimbalyst

There’s a million AI news articles, resources, tools, and fundraises every week. You can’t keep track of everything. Here’s what mattered - the 1 big story and key news.

Nano Banana 2 Changes the Math on AI Image Generation

Remember when Nano Banana went viral last August? Then Nano Banana Pro dropped in November with studio-quality output. Now Google shipped Nano Banana 2. And it combines the quality of Pro with the speed of Flash:

That makes it the biggest drop of the last week. The real story is three things that actually matter for builders:

1. Text in images finally works

If you’ve ever tried putting text inside an AI-generated image, you know how painful it was. Blurry, misspelled, wrong fonts. Every time. Nano Banana 2 can generate posters, banners, and infographics where the text is actually readable. It even handles translation across languages without breaking the layout. If you’re building anything customer-facing with AI visuals, this was the #1 blocker. It’s gone now.

2. The iteration speed is completely different

You can now test a bunch of variations quickly at low resolution, pick the one that works, and scale it up to 4K. The feedback loop that used to take minutes now takes seconds. For anyone doing product mockups, marketing assets, or even just exploring ideas visually, this changes how you work.

3. The API cost dropped 4x

It also pulls real-time web knowledge from Gemini during generation. So it can reference current events and real brand elements instead of hallucinating something from 2023. Surprisingly useful for marketing use cases.

If you read my Nano Banana Pro prompts, those all still work. Just better and faster now. The API is live. Go test it.

Key News

• Anthropic told the Pentagon no. The Department of Defense wanted Claude’s safety limits removed. Anthropic walked away from a $200M contract. And it got labelled a “Supply Chain Risk” as a result. That hurts their enterprise deals, but hasn’t stopped them going from $14B to $19B ARR in about a month. It is one of the most effective consumer brand strategies in AI.

• Perplexity is now integrated into all Samsung Galaxy S26 phones. Their Sonar API powers Bixby’s search. Samsung’s own assistant now runs on Perplexity’s infrastructure.

• DeepSeek’s upcoming V4 model was trained on smuggled Blackwell chips. The DeepSeek narrative was always “2,000 H800 chips, $5.6 million, and Chinese ingenuity beat $100 billion in American spending.” That story just collapsed.

Tool Updates

• Perplexity launched Computer. Think of it as “OpenClaw for the less technical.” It orchestrates 19 AI models to complete entire workflows. You describe what you want done, it breaks it into subtasks, picks the best model for each one (Claude for reasoning, Gemini for research, Nano Banana for images, Grok for speed), and runs in the background.

• Cowork got Scheduled Tasks. Claude stopped being software you talk to and became software that works while you sleep.

Fundraising

• WorkOS raised $100M at a $2B valuation. They sell SSO and directory sync. Sounds boring until you see the customer list: OpenAI, Anthropic, xAI, Cursor, Perplexity, Replit, Vercel. Every company building next-gen enterprise software runs on WorkOS. That’s a bet on the entire AI application layer.

• OpenAI raised $110B. The real signal is in two technical terms most people scrolled right past. Microsoft gets “stateless API.” Amazon gets “stateful runtime environment.” OpenAI just handed the growth segment to Amazon while leaving Microsoft with the architecture that’s becoming the commodity layer.

I Open-Sourced Claude Code Setup and 5 Key Skills

I pay for Google AI for Nano Banana + NotebookLM. Gemini itself is still much worse than Claude.

That’s because Claude Cowork and Claude Code are absolute powerhouses for any knowledge worker.

But the problem everyone runs into is starting with Claude is starting. So I’m open sourcing a free ‘0 to 100’ guide with 2 new Github repos today.

Deep Dive Outline

I’ve put together the ultimate startup guide for Claude Code and Cowork:

1. Building Claude’s Memory

2. 5 Starter Skills

3. Claude Code Setup

4. Where to Go From Here

Subscribe now

1. Building Claude’s Memory

So let’s start with the new Import Memory feature:

You go to Settings > Capabilities, find the Memory section, and select “Start import.”

It has you put this prompt into your existing AI (Google or OpenAI likely):

I’m moving to another service and need to export my data. List every memory you have stored about me, as well as any context you’ve learned about me from past conversations. Output everything in a single code block so I can easily copy it. Format each entry as: [date saved, if available] - memory content. Make sure to cover all of the following — preserve my words verbatim where possible: Instructions I’ve given you about how to respond (tone, format, style, ‘always do X’, ‘never do Y’). Personal details: name, location, job, family, interests. Projects, goals, and recurring topics. Tools, languages, and frameworks I use. Preferences and corrections I’ve made to your behavior. Any other stored context not covered above. Do not summarize, group, or omit any entries. After the code block, confirm whether that is the complete set or if any remain.

Then you copy the results into Claude’s memory import feature (not a regular chat), and Claude will update its memory from your prior AI.

If you haven’t developed a memory with AI, open a chat with Claude and complete the following prompt:

I'm [name]

My most important goals are: [1. X, 2. Y]
I will be using Claude to [X, Y, X] 
My profession is [X]
My biggest roadblock at work is [X]
One thing I want to get better at is [X]

Add this information to your memory about me so you can be my personalized AI copilot that 10x's my ability to reach my goals with the goal of model obsolescence and self-sufficiency.

This might seem like a small thing, but it makes a huge difference to the quality of the responses you receive.

2. 5 Starter Skills

The most important feature that Claude has is skills. This is the big differentiator from ChatGPT and Google.

ChatGPT Projects and Gemini Gems are just wrappers on top of their AI. Skills are how you actually train an AI to do work for you.

But most people struggle with starting with skills. So steal mine:

Get the Claude Skills