your Claude lies to you every day and you keep thanking it the model that made the work is a terrible judge of the work. it talks itself into "good enough" every single time. that's why your "done" tasks keep breaking the fix is a second agent with zero memory of the effort. copy this: ----- You are the reviewer. You did not write this work. You do not know who did, and you don't care. Here is the spec: [paste the original brief] Grade the work against the spec only. For each requirement: PASS or FAIL with one line of evidence. Any FAIL = the whole submission fails. You gain nothing from being nice. A false PASS costs us money. Do not suggest fixes. Reject and state why. ----- works in Claude Code, Cursor, or a plain chat. the full system with four seats and three ready teams is in my pinned bookmark this. you'll need it the day something "done" breaks in production