Reviewing AI-Written Code

If the model wrote it and the model reviews it, you have one opinion twice.

Ask the session that just spent forty minutes building something whether it is correct, and it will tell you yes. It is not lying to you. It has no memory of deciding anything, so it reads its own output the way it reads any other code: as a plausible pattern.

The five things it gets wrong, reliably:

Hardcoded values it invented because it could not see your config,
Catch blocks that swallow the error and return an empty array,
Access rules that pass every test because your tests have one user,
Logic it duplicated because the original was outside its context window,
Edge cases handled with total confidence and roughly one in four wrong 🔥

The fix is not a better review prompt. It is a different reviewer.

Fresh session. Paste the diff with no history of how it was written. Tell it to argue the change should not be merged. A model with no memory of authoring the code has no position to defend.

I wrote the whole thing up. Eight review prompts, real before-and-after code, and a checklist to run before you merge.

Free. No email wall.

Comment "REVIEW" 👇 and I'll send it to your DMs.