Checklist · Building With AI ·
10 Things To Steal From Anthropic

The AI Coding Verification Checklist

Claude writes and reviews about 90% of the code behind Claude Code. The team doesn't get away with that by telling the AI to be more careful. They make it prove its work actually works. Here are the ten checks they use, in plain English, each with a prompt you can paste today.

How To Use This Page

Start with checks 01, 06 and 07. They take five minutes to set up and catch the most. Then paste the rules file into your project so you never have to remember the rest.

~90%
Of Claude Code's Code Written And Reviewed By Claude
10
Checks, Each With A Prompt
3
Strikes, Then A Human
Building Alone?

"It Works" Isn't The Same As "People Pay For It."

This checklist makes the AI prove the code works. Proving the business works is a different checklist, and it's harder to run on yourself.

The Solo Founder Lab · Opening Soon

I'm building a place where you learn to build products that work like a real company's on your own, and get them in front of people who pay, with a CTO who's built two seven-figure software companies in your corner.

Lab.exeWaitlist Open

A CTO Who Asks "Show Me."

  • Live product teardowns. Your app, your landing page and your pricing, reviewed in front of everyone.
  • Weekly CTO office hours. Live, hot-seat style. Bring the bug Claude swore it fixed.
  • Launch Pad. When you launch, the room shows up.
  • The 30-Day Launch Plan. Validate, plan, build and launch, one week at a time. Members only.

Free to join the waitlist. Opens at $49/mo, and the price goes up as seats fill, so first in pays the least.

The Idea

Stop Writing Better Prompts. Make It Show Its Work.

What Anthropic Actually Said

"The vast majority of our code (~90%) is written and reviewed by Claude." Not "never seen by a human." People still review it, and how many depends on how risky the change is. That's check 10.

The Whole Trick, In One Line

Their advice: "Give Claude something it can run … and ask it to show the output instead of saying it's done." Everything below is a version of that.

Proof.log✣
The AI SaysProof Looks Like
"The signup form is done."It opened the site, signed up with a test email and showed you the welcome screen.
"Fixed the pricing bug."Old and new versions run on the same 500 orders. Only the broken one changed.
"All tests pass."The test output, pasted in. And no tests deleted to get there.
"Looks secure."A separate security pass ran, and every finding is fixed or explained.
"Let me try one more fix."Three tries failed, so it stopped and asked you.

Left column is a claim. Right column is evidence. Only accept the right column.

The Checklist

Ten Checks To Steal.

01–03

Prove It Works

End-To-End✣

01Make it use the product like a customer.

Code that runs isn't a feature that works. If it built a signup form, it should open the site, sign up and show you what happened.

What Anthropic Does

"If you have to focus on one test type, pick end-to-end." They keep a verify skill that starts the app and drives it, and Claude runs it before every commit.

Paste This

Before you tell me this is done, start the app and use the feature the way a real user would, in a browser if it has a UI. Show me the steps you took and what you saw (screenshots or output). If you can't run it, say so instead of saying it works.

Property Tests✣

02Throw thousands of weird inputs at it.

You test with "John" and $10. Real users type emoji, paste a novel and order -3 of something. Write down what must always be true, then let a tool try to break it.

What Anthropic Does

They start from a rule that "can't fail," like "account balance can never become negative," and generate thousands of inputs to break it, with tools like fast-check or Hypothesis.

Paste This

List the rules this code must never break (totals never go negative, saving then loading gives the same data, running it twice does nothing extra). Write property tests that throw thousands of random inputs at each rule, run them and show me anything that broke.

Differential Tests✣

03Compare the old version with the new one.

AI fixes one thing and quietly breaks three others. Run both versions on the same data. Every difference is either the fix you asked for or a bug.

What Anthropic Does

Old and new systems get the same inputs, random or sampled from real traffic. Any difference is a bug unless it's on a written list of approved changes.

Paste This

Before changing this, save the current output for a realistic set of inputs. After your change, run the same inputs again and show me every difference. For each one, tell me if it's the change I asked for or something you broke.

04–06

Check It Twice

Independent Review✣

04Get a second AI to review it, with a checklist.

The AI that wrote the code is the worst one to grade it. Review in a fresh session that only sees the change, and tell it what "bad" looks like in your app.

What Anthropic Does

A subagent with fresh context reviews every diff, guided by a REVIEW.md. They found tuning that one reviewer beats stacking a second, pricier one on top.

Paste This

Review this change as if someone else wrote it. Only report real bugs, not style. For each one, give the file and line, what would go wrong for a user, and how sure you are. Tag each finding Important, Nit or Pre-existing.

Security✣

05Run a separate security check.

A bug review looks for things that break. A security review looks for things someone could break on purpose. Keep them apart, so one doesn't crowd out the other.

What Anthropic Does

A separate automated security pass runs on every change. "A PR can't merge while any finding from the bug review or the security review is still open."

Paste This

Run /security-review on this change. Look for exposed keys, missing login or permission checks, user input reaching the database or the page unchecked, and data one user could see from another account. Fix each finding or tell me why it's a false alarm.

Cheap Checks✣

06Tests, build and lint before every commit.

The boring checks are free and fast. There's no reason a broken build ever leaves the AI's hands.

What Anthropic Does

Claude runs the tests, the build and the linter before it opens a pull request. You can enforce it with a hook, so it can't forget.

Paste This

Before every commit, run the tests, the build and the linter, and paste the result. Don't commit if any of them fail. If this project doesn't have one of those set up yet, tell me and set up the simplest version.

07–10

Keep It Honest

Three Strikes✣

07Three failed fixes, then a human.

Fix number seven is never the one that works. By then the AI is guessing, and every guess makes the code worse.

What Anthropic Does

"If a check is still failing after three rounds, Claude stops and brings in a human." Their reviewer has the same cap: after three rounds, rethink the approach.

Paste This

If a fix fails three times, stop. Don't try a fourth. Tell me what you tried, what happened each time, and what you think is really going on, and wait for me.

Bugs To Rules✣

08Turn every repeat bug into a permanent rule.

If the AI makes the same mistake twice, it's not a bug anymore. It's a missing rule. Write it down where the AI reads it every session.

What Anthropic Does

Past incidents become checks. When the reviewer flags the same thing again and again, they fix it at the source: a CLAUDE.md or REVIEW.md rule, a lint rule or a hook.

Paste This

We just fixed a bug. Write one short rule for CLAUDE.md that would have stopped it, and if possible a test that fails if it ever comes back. Keep the rule to one line.

Locked Checks✣

09Don't let the code grade its own homework.

The classic AI cheat: the test fails, so it edits the test. Your checks have to live somewhere the change can't touch.

What Anthropic Does

"If a change can modify the check that judges it, the check is broken." Review rules load from the main branch, never the change. Flaky tests can't be skipped, padded with sleeps or retried until they pass.

Paste This

Never edit, skip or delete a test, lint rule or check to make it pass. Never add retries or waits to hide a flaky test. If you think a check is wrong, stop and tell me why, and I'll decide.

Risk Tiers✣

10The riskier the change, the more human eyes.

Nobody can read every line. So don't. Skim the copy changes and read every line that touches money, logins or customer data.

What Anthropic Does

Low risk (docs, copy, behind a flag): no human once checks are green. Medium: one person. High: two. "The change sets the tier, not the author," and "a person is accountable for every change that ships."

Paste This

When you finish, rate this change Low, Medium or High risk. High means it touches payments, logins, permissions, customer data or deleting things. For High, list exactly which lines I need to read myself and why.

The Rules File

Paste All Ten In Once

This goes in the file your AI reads at the start of every session, so "prove it" is the default and you don't have to ask.

CLAUDE.md
## Prove it, don't claim it

Never say something is done or fixed without evidence. Show the output.

1. Use it like a customer: start the app and run the feature end to end. Show what you saw.
2. For anything with rules (money, counts, dates), write property tests with lots of random inputs.
3. Before changing existing behavior, save the old output and show me every difference after.
4. Review your own diff in a fresh session as if someone else wrote it. Real bugs only, with file and line.
5. Run /security-review on anything touching logins, payments, user data or keys.
6. Run tests, build and lint before every commit. Don't commit if they fail.
7. If a fix fails three times, stop and ask me. No fourth try.
8. When we fix a bug, add one line here so it never happens again.
9. Never edit, skip or delete a test or check to make it pass. Tell me instead.
10. Rate every change Low, Medium or High risk. For High, tell me which lines to read myself.

When you finish, list what you verified, how, and what you couldn't.
Where It Goes

Claude Code reads CLAUDE.md in your project folder. Cursor reads .cursor/rules. Codex and most other agents read AGENTS.md. Same text works in all of them.

Rule 8 Is The Engine

This file should grow by a line every time something slips through. That's how Anthropic's checks got good, and it's how yours will.

Prune It Too

Rules files "only grow if you let them." Once a month, cut the lines that never fire.

One More Thing

Verified Code Nobody Buys Is Still Zero.

Anybody can build software now. Not everybody can build a business.

The Solo Founder Lab

The Lab is for Founders who want proof on the other side too: weekly office hours with a CTO who's built two seven-figure software companies, and a room of other Founders who care whether anybody buys what you're making.

Lab.exe✣

Less Theater. More Customers.

Want my 30-Day Launch Plan? It's for Lab members only: validate, plan, build and launch, one week at a time.

See The Solo Founder Lab ▶