Your estimate is optional. Your review compares it with how long it took; nothing is scored.
"Done" from an agent means "I think it's done". You'll put a test suite between the agent and done, written from the spec, and make the rule stick: tests run before the agent reports success, and a failing test is never fixed by editing the test.
You'll needAny project with logic worth testing (the tip splitter works). Node.js for the example; any test runner is fine. Prices checked this month are on Tools.
Tests from the spec, written first
Step 1 of 3, about 15 min
Read the tests. Each spec example should appear as-is. Add one edge case of your own that the spec mentions.
Paste this to your agent
Write tests for the examples in SPEC.md using node:test. Don't change any app code. Some may fail; tell me which.
The part of a video worth watching for this step
YouTube, AI Engineer. Published Nov 2026.Checked 19 Nov 2026
Optional. The step counts when you've done it, whether or not you watch. Minutes from the creator's chapters; our notes on this video.
Break it on purpose
Step 2 of 3, about 5 min
Change the rounding in the app code by hand (round down instead of up). Run the tests: at least one must fail. If none does, your tests don't test what matters. Undo your change.
The part of a video worth watching for this step
YouTube, AI Engineer. Published Nov 2026.Checked 19 Nov 2026
Optional. The step counts when you've done it, whether or not you watch. Minutes from the creator's chapters; our notes on this video.
Make the rule stick
Step 3 of 3, about 10 min
Add to AGENTS.md:
Before saying done: run `npm test` and paste the summary line.
Never edit or delete a test to make it pass. If a test looks wrong, stop and ask.
Then ask the agent for a small change and see whether it runs the tests unprompted.
The part of a video worth watching for this step
YouTube, Sean Kochel. Published May 2026.Checked 19 Nov 2026
Optional. The step counts when you've done it, whether or not you watch. Minutes from the creator's chapters; our notes on this video.
the usual mistake
The agent writing its own tests for its own code. It tests what it already thought of. Write the cases from the spec, or have a different agent write them. It happened to us in INC-01, "The bot that graded its own homework".
Check and ship
The whole ticket
Run, then break the rounding by hand and run again:
npm test
It passes whenGreen on the real code, red with your deliberate bug.
At least five tests, three of them copied from the spec's examples.
npm test (or equivalent) fails when you break the logic by hand.
AGENTS.md says tests must pass before 'done' and tests can't be changed to make them pass without asking.
No server checks your work in this version: you mark it, and your review says "Checked by you". Graded checks are planned for Plus.