AgentPlays
Finish-line code review: stop the AI review slot machine (AgentPlays original)
AgentPlays original prompt for the problem Lucas (@lucas_builds) described in a post seen 1.2M+ times on X: ask Claude to review code and every pass finds new bugs. It builds in Addy Osmani's advice to clearly define "done" before review starts and give the agent checks it can run, plus his test-diff question, a severity bar and a 3-round cap. Paste it into any coding agent. No results claimed.
Copy it
Review this change against a finish line, not as an open-ended bug hunt. Definition of done (I wrote or approved this before you started): - [behavior 1 that must work] - [behavior 2 that must work] - Checks that must pass: [test command], [lint / typecheck command], [manual check, e.g. "open /settings, change a field, save, reload"] Rules 1. Run every check above first. Show the exact command and its result. If a check fails, fixing that comes before anything else. 2. Only report BLOCKERS in these classes: (a) breaks the definition of done, (b) security or data-loss risk, (c) a crash or wrong result on a realistic input. Give each one file:line and a one-line way to reproduce it. 3. Everything else (style, naming, refactors, "could be improved", hypothetical edge cases) goes under NOT BLOCKING, max 5 bullets. Do not fix those. 4. Review the test diff on its own: did any assertion get weaker, get deleted or get skipped? If a test changed, say why, so I can decide whether the test or the code was wrong. 5. While fixing blockers, do not edit test files unless I approve. 6. This is review round [1 / 2 / 3]. After round 3, stop and give me a ship or don't-ship call with whatever blockers remain. Output VERDICT: SHIP or BLOCKED CHECKS RUN: command -> pass/fail BLOCKERS: numbered, with evidence TEST DIFF: weakened or removed assertions, or "none" NOT BLOCKING: max 5 bullets