Nine seats advise. The tenth decides.
A disciplined, nine-seat review board you can run with one person and one language model — built to find the defects that a green CI cannot see.
You have tried this. You asked a model to review your code and got a confident, well-formatted page of concerns. Some were real. Most were not. You had no way to tell which, and after chasing two or three plausible-sounding nothings you quietly stopped running it.
That is not a model limitation. It is a missing process.
Real review boards work because participants cannot opt out of the structure: seats with narrow mandates, a format that findings must fit, a chair who adjudicates rather than averages, and a hard rule about what is allowed to block a merge. Strip those out and you have a meeting. Strip them out of an AI review and you have theater.
Everything published here is about putting that structure back.
Why generic review prompts return noise, the one rule that filters it, and a complete copy-paste prompt with a worked example.
A working ToR you can adapt, the right-sizing table that stops the board becoming a tax, and the five ways review boards fail.
The kit
Eight laws of doctrine, each with the failure that produced it. Eleven persona charters as paste-ready system prompts. Five orchestration patterns. A sixteen-pattern probe catalogue. Two fully worked reviews. Two fill-in worksheets. A 21-page guide.
Distilled from months of daily use on a 200,000-line trading platform, where the board reviewed every merge and repeatedly caught defects the test suite passed.