Context
Most code review with AI is ad-hoc. A reviewer pastes a diff into Claude, asks "any issues?", reads the answer, moves on. It works once. It does not scale.
This is the SOP we run weekly. It's boring. That's the point.
The SOP
- Author opens the PR with a 3-line description: what changed, why, and what to look for.
- Author runs the self-review prompt (below) and pastes the output into the PR description under a
## AI self-review heading.
- Reviewer runs the reviewer prompt against the diff, reads the output, and uses it as a starting point for their own pass.
- Reviewer leaves human comments. The AI output is never the final word.
The prompts
You are reviewing a pull request. The diff is below.
Report in this format:
- Correctness risks (concrete, by file/line)
- Test coverage gaps
- Style or naming notes (only if non-trivial)
- Questions for the author
If the diff is small or low-risk, say so and stop.
What changed after 6 weeks
- Author self-review caught 30% of issues before reviewer saw the PR.
- Reviewer time dropped from 25 min average to 14 min.
- We killed the "Style or naming notes" section after week 3 because it generated noise. Now we omit it.
Pitfalls
- Don't skip the human pass. AI misses business logic.
- Don't paste secrets into the prompt. Use a local model or scrub first.
Variations by scale
- Solo: skip the reviewer step, run the self-review prompt before every commit.
- Small team (2–10): as written.
- Mid (10–50): add a per-language prompt variant, store in a shared repo.
- Enterprise: route through an internal proxy, log prompts for compliance.
Comments