Start with the actual diff and its risk
Describe the user-visible change in one sentence. Compare every changed file with the requested scope; check generated files, dependencies, CI, environment settings and deleted checks. A summary written by the agent explains its intention, while the diff shows what it did.
A copy edit may need a visual check. A component needs a build, focused tests and the changed user flow. Authentication, payment and personal-data changes deserve independent review and a rollback plan.
- Are any files or refactors outside the agreed scope?
- Can you explain the changed behaviour without relying on the agent summary?
- Were permissions, dependencies or deployment settings changed?
A build passed while the answer was wrong
An earlier DottorLone quotation example listed €85 per month for 12 months plus a €300 setup fee, but published €1,620 as the total. The correct total is €1,320. That error reversed the comparison, even though the static page and build passed.
The repair corrected the editorial example and added a calculator test using integer cents. The important review question was not “does it compile?” but “do these inputs produce the right total and conclusion?”
const totalInCents = 8500 * 12 + 30000;
assert.equal(totalInCents, 132000);
// Also check different durations, missing inputs and the result in the browser.Run the checks that match the change
- Run the project’s documented production build and relevant automated tests.
- Exercise the changed user flow, including empty and error states.
- Check desktop and a narrow mobile viewport, keyboard focus and the browser console when the UI changed.
- Inspect inputs, authorisation, logs and frontend bundles for sensitive data when risk warrants it.
- Record commands, results, and anything you could not verify.
Task: <USER OUTCOME>
Expected files: <LIST>
Actual files: <LIST>
Build: <COMMAND AND RESULT>
Focused test: <COMMAND AND RESULT>
Manual flow: <STEPS AND RESULT>
Unverified: <LIMITS>
Decision: <ACCEPT / REVISE / SECOND REVIEW>Write a finding another person can fix
Give the exact condition, actual result, expected result, impact and a reproducible check. “The math seems wrong” gives less help than the €85 × 12 + €300 example above. Link the finding to the changed line or user path and separate confirmed failures from hypotheses.
Do not merge merely because an AI reviewer and CI both say green. Their checks cover different questions, and neither replaces the human decision about the task.