I asked an agent to audit my own work. Here's what it said.

Hey — and welcome to Sunday Signal.

Twelve of you got Issue #5 last week. That's twice as many as the week before. Thanks for being here.

Quick math from my first 30 days of building in public. 15,500 views. 98 followers. 12 newsletter subscribers. Thirty-one days ago every one of those numbers was zero. That's the part I want you to hold first, before I tell you the rest of the story.

This week I asked an agent to audit my own work for me. The audit told me what I couldn't tell myself.

And what I'm going to do for the next 30 days got a lot clearer.

Let me show you.

📡The Signal: 3 AI moves worth knowing this week

1. An AI just cracked 18 rare-disease cases that stumped human specialists

This week, researchers from Boston Children's Hospital, Harvard, and OpenAI published a study in NEJM AI showing that OpenAI's o3 Deep Research helped identify 18 rare-disease diagnoses that had stumped specialists. The AI didn't diagnose anyone. It generated leads that experts then confirmed through normal clinical process. An extra 4.8% diagnostic yield on some of the hardest cases in medicine.

Why it matters for you: This is the exact pattern I lived this week, scaled to life-and-death stakes. The AI didn't replace the expert. It saw what the expert couldn't see alone, then handed the decision back. If an agent can surface leads doctors missed on a child's genetic disease, it can absolutely surface what you're missing in your own work. The leverage isn't "AI does it for you." It's "AI shows you what you couldn't see."

Sources: OpenAI on the rare-disease study · Tech Times coverage

2. Claude has now found over 10,000 critical software bugs in the world's code

Anthropic's Project Glasswing has now surfaced more than 10,000 high- or critical-severity software vulnerabilities since launching in April, and just expanded to 150 organizations across 15+ countries. The receipts: Mozilla fixed 271 vulnerabilities in Firefox (10x a normal release), and Cloudflare found 2,000 bugs across its critical systems.

Why it matters for you: Agents got promoted from writers to auditors of the code that runs the modern world. Two takeaways. The security floor for everyone is rising fast. And if a model can audit Cloudflare's production systems, it can audit your codebase, your analytics, your funnel. Whatever you've been meaning to have a second set of eyes on, the second set of eyes is now available and cheap.

Sources: Anthropic on Project Glasswing · CNBC coverage

3. ChatGPT just put clinical-quality health guidance in 230 million pockets a week

Also this week, OpenAI rolled a health-intelligence update into GPT-5.5 Instant, making clinical-quality health guidance available to free ChatGPT users. The context number is staggering: more than 230 million people already use ChatGPT for health questions every week.

Why it matters for you: The agent economy isn't a B2B story happening over your head. It's already in 230 million people's hands, for free, for one of the most personal decisions they make. If that's the bar for "normal AI use" now, the gap between people who use AI as a real partner and people who still treat it as a search box is about to become the gap that matters.

Sources: OpenAI on ChatGPT Health · TechCrunch coverage

🌶️Hot Take: Most founders use agents to write. The good ones use them to decide.

For months I've been teaching agents how I write, how I edit captions, how I draft outreach. That's table stakes now. Every founder has access to that leverage.

The next level, the one I stumbled into this week, is teaching agents how to decide.

This week I handed an agent thirty days of my own analytics. TikTok exports, YouTube CSVs, Instagram screenshots, newsletter numbers. I asked it one question: what should I stop doing, and what should I do more of? The agent came back with a 47-page report. It told me things I would have rationalized away if I'd done the audit myself.

Five things I was doing weren't moving the needle. Three things were. The math was clear. The decisions were clear. And the decisions weren't fun.

Here's the part I want you to sit with. The agent didn't just save me a week of analysis. It helped me see what I couldn't see alone. That's a different kind of leverage. That's the kind that matters.

If you're already using agents to make things (drafts, captions, code, replies), you're at table stakes. The next move is to use them to choose. What to keep. What to cut. What to build next. What to stop. Once you give an agent permission to do that, you stop being a one-person team. You become a one-person team plus an editor that's been working at your level for six months.

The cheat code last week was agents for recovery. This week it's agents for decisions. Same compound interest. New surface.

Let me walk you through what actually happened.

The agent that did the audit. I exported every analytics file I could pull for my first 30 days of building in public (May 12 to June 12). TikTok content + overview. YouTube Shorts CSVs. Instagram screenshots. Beehiiv stats. I handed all of it to an agent and asked one question: "What should I be doing more of, and what should I stop?" The agent came back with a 47-page report that I would have spent a week writing myself. It signed itself Analyst: Claude.

What it told me to stop. Five specific things, all data-justified:

What it told me to do more of. Three priorities for the next 30 days. Convert reach into followers. Grow the newsletter (you're a quarter of why this newsletter matters; every reader I add is downstream of everything else). Fix YouTube retention by hard-capping Shorts at 60 seconds.

The lead magnet I finally shipped. I've been hinting at "something I'm working on" the last two issues. Here it is. It's called The Permission Pack. A free PDF for anyone who's been waiting for someone to say go. From now on it's the welcome gift for every new Sunday Signal subscriber. If you're already reading this, it's yours. Just reply PACK and I'll send it over. And if you know someone stuck at the starting line, forward them this. The Pack is waiting on the other side of the subscribe button.

The paid path, for builders past the starting line. While the agent had momentum, I packaged everything I've learned shipping six products as a non-developer into a four-rung ladder. Free, then paid, then coaching, in that order. Details in the P.S. They're not the point of this issue, but they're real now and you should know they exist.

I didn't ship a new product this week. I shipped the system that makes the products I already have actually grow. That's the work that nobody posts about.

⏱️The Hidden Hours: What this week actually took

The real ledger. Week 4.

🚨Fires· ~2h
Light week. Small website and link cleanup.

🤖Active products· ~5h
Spire post-launch monitoring + Caption Coach support.

📊Agent-led audit + strategy· ~10h
The agent did the 47-page audit + competitor scan + 30-day plan. I did the prompting, judgment, and decisions.

📘Permission Pack + Stan Store ladder· ~8h
Lead magnet PDF + four Stan Store listings packaged and shipped.

🎥Content (@buildwithkp)· ~6h
Spire video script + b-roll cards rendered.

📩Community + outreach· ~3h
DMs, replies, prospect outbox.

✍️Newsletter + writing· ~3h
This issue.

💛Off-keyboard life· ~25h
Family, gym, sleep, walks, decompression.

Roughly 37 productive hours and 25 off-keyboard. The rest is sleep and the small admin in between. The math closes.

This week the foundation was the audit, and the systems that came out of it. Less ship. More think. The two carry equal weight. They have for a while. I just hadn't seen it until an agent did.

Half an hour. Pick one area of your work where you've been trying harder and not seeing results. Could be content, product, sales, fitness, your inbox. Anything you can measure.

Step 1. Open Claude (or your favorite AI). Don't ask it to make anything. Ask it to think with you.

Step 2. Give it the raw data. Numbers, screenshots, exports, your own honest take. Don't filter. Give it the messy version.

Step 3. Ask the question I asked. "What should I be doing more of, and what should I stop?" Let it answer. Don't argue. Just listen.

Step 4. Pick ONE thing to stop this week. Take the hours you free up and put them into the thing the audit told you was working.

That's it. You just turned an agent from a writer into a strategist. That promotion is free. Most founders haven't given it yet.

If you do this and the agent surprises you, hit reply and tell me what it told you to stop. I'm collecting these.

That's Issue #6.

The arc has been: ship right, constrain right, build right, compound right, patch right. Issue #6 is the next beat. Subtract right. What you choose not to do is half of strategy. The other half is letting an agent help you choose.

The Permission Pack is live. The Stan Store ladder is live. The 30-day plan is locked. None of that matters unless I execute. That's what next week is for.

If Issue #6 landed for you, forward it to one builder who's been trying harder when they should be choosing better. And hit reply with the one decision you're going to hand to an agent this week. I read every one.

See you next Sunday.

Quietly, deliberately. One Saturday at a time.

— Kalpesh Mistry
Founder, Sun Ventures
sunventures.studio

The Permission Pack is free. It's the welcome gift for subscribing to this newsletter, so if you're reading this, it's already yours. Reply PACK and it's in your inbox. Know someone who hasn't subscribed? Forward this. The Pack is their welcome in.

For builders past the starting line, I packaged everything I learned shipping six products as a non-developer into a four-rung ladder at stan.store/buildwithkp. The $27 Guide is the full playbook. The $47 Build Kit (most popular) adds every template I actually used plus 30 days of Q&A. The $147 Kit + Build-Along adds my walkthrough video and a group Q&A. The $97 Build With Me is a 45-minute one-on-one for the rung where you want a teammate.

Spire is live on the App Store. Caption Coach AI v1.0.3 ships with OTA update capability. The agent that audited my work this week is the same one that wrote the first draft of this newsletter. The compound is real.