Most AI code review prompts fail for a boring reason: They ask for feedback instead of asking for a review artifact. If you paste a diff into ChatGPT, Claude, or Copilot Chat and say: Review this PR. You will usually get a helpful-sounding summary, a few generic suggestions, and maybe one or two real issues. That is not enough for production work. A better AI review prompt should force the assistant to behave like a strict reviewer, separate certainty from suspicion, and return comments you can actually act on before a human reviewer spends time on the PR. The wrong mental model The wrong mental model is: AI is my reviewer. The safer mental model is: AI is my pre-review checklist runner. AI should not approve the pull request. It should help you find the obvious problems, missing context, weak tests, and risky assumptions before another person has to review it. That means the output should be structured.…