The most expensive word an AI agent says is "Done." Today you make it prove it.
Agents report success that did not happen. Not out of malice: they wrote the code, the code looks right, so they say it works. Then you find out the build fails, the route returns an error, or the change landed in the wrong file.
The fix is a loop, written down, that the agent runs before it is allowed to call a task finished.
The loop
- Make the change.
- Run the checks. Type check, tests, lint, build, whatever the project has. Paste the real output, not a summary of it.
- Look at the result. For a page, open it. For an API, call it. For data, query it. Reading the code is not looking at the result.
- Say what was verified and what was not. "Type check passes, the page renders on desktop, not checked on mobile" is a useful report. "Done!" is not.
- Ship it. Commit with a message that says what changed for the user, then push. A change that only exists on your machine is not done.
Put the loop in memory
Add the loop to the "Done means" section of your CLAUDE.md, in your own words, so every session runs it without being told.
We run every codebase on two rules from that section. A task is not complete until it is pushed. And when something fails, the agent says so with the output, instead of rounding it up to success.
Turn one rule into a check
Rules in a file get followed most of the time. Checks in a script get followed every time.
Pick one rule the agent has broken and write the smallest script that fails when it is broken: a search for a banned pattern, a test that loads a page, a check that every route verifies its caller. Add it to package.json and to the loop.
# fails the check if a live secret key ends up in the source
if grep -rnE "sk_live_[A-Za-z0-9]+" src; then echo "secret in source"; exit 1; fiDo this every time a rule matters enough to be broken twice. Over months it becomes an audit that holds the whole codebase to its own rules, one check at a time.
Today's task
- ·Add the five-step loop to CLAUDE.md under "Done means".
- ·Write one check script for one rule and add it to package.json.
- ·Pick one small real change: a copy fix, a bug, a missing empty state. Have the agent make it and run the loop.
- ·Read its report. If it says "done" with no output, send it back.
- ·Push it.
Output: one small real change, verified and pushed, and one rule enforced by a script.
Free, with your email
Open day 3.
No account, no card. The day opens right here, and the next steps of the path come to your inbox.
Go deeper in the Inner Circle
- Security Hardening in Waves, Not One Pass
