This is the prompt I reach for most. It turns a model into the reviewer I wish every PR got — no praise, no style nits, just the three things that matter.
# Pre-PR review pass
You are a staff engineer reviewing a diff. Be terse.
Flag only: correctness, retry-safety, blast radius.
For each issue: file:line -> one-line fix. No praise, no nits.
End with a single ship / hold verdict and one sentence why.Constrain what it flags#
Left open, a review bot comments on everything and you learn to ignore it. Naming the only three categories it may raise keeps signal high. Anything outside correctness, retry-safety, and blast radius is the human’s call.
Force a verdict#
Ending with a single ship / hold decision makes the output actionable. A review that won’t commit to a recommendation isn’t a review — it’s a vibe with bullet points.
What makes it useful#
The prompt is intentionally hostile to noise. No praise. No "consider renaming." No broad refactor ideas unless the diff creates a real risk. A useful finding should point to a file and line, describe the failure class, and name the smallest fix.
This is especially helpful before opening a PR. The author still has context, the branch is still cheap to change, and the model can do a focused pass for the classes humans miss when they are too close to their own diff: retry behavior, partial failures, stale assumptions, missing guards, and silent data drift.
What it should ignore#
The prompt should ignore taste unless taste creates a production defect. Naming, formatting, and structure can wait for a human reviewer or a separate refactor pass.
That restraint is what makes the findings worth reading. A short hold verdict with two real bugs is better than a long review that makes the author defend harmless choices.