Settled into a workflow that keeps AI assistance useful instead of noisy: I write the first draft of any non-trivial change myself, then ask the model for a review with a narrow brief — "find correctness bugs and race conditions, ignore style". Broad prompts produce style noise I start ignoring; narrow prompts produce findings I act on.
Two rules that made it stick:
- The model reviews, I decide. Every suggestion gets accepted or rejected with a reason I could say out loud. Rubber-stamping diffs I did not fully read is how subtle bugs ship.
- Context in, secrets out. I paste the diff and the relevant types, never env files or production data.
Net effect after three months: fewer review round-trips with teammates, and my self-review catches edge cases I used to delegate to "QA will find it".