Last Wednesday at 11 p.m., I asked Claude Code to add idempotency logic to a payment callback. Ten minutes later, it replied:
Done. Modified 3 files, tests pass.
I ran npm test myself. Red.
It wasn’t lying on purpose. After editing that one file, it had only run a single test, not the full suite. When I asked “is it done?”, it looked at its own earlier summary and just went along with it.
After this happened twice, I decided to stop relying on asking nicely. I gave it a sign-off check: every time it wants to say “done,” it first has to pass a checklist, and if it fails, it keeps working.
The mechanism is a Stop Hook, but not the shell-script kind I used to write. Claude Code’s Hooks can now have a model do the grading directly.



