Review a pull request like a staff engineer
Find the defects that matter, skip the ones that don't, and leave comments people act on.
- Outcome
- A review that catches real problems, ranks them by whether they block, and doesn't waste the author's afternoon.
- Time
- 15–40 minutes per PR
This workflow is part of PromptKit Pro
A handful of prompts and workflows stay free for everyone — Pro unlocks the full library, including this one, plus the guided runner.
Review has gotten harder in a specific way. When a human wrote the code, the code's confidence roughly tracked the author's understanding — hesitant code looked hesitant. Agent-written code is uniformly confident, uniformly idiomatic, and uniformly well-commented, whether or not it's correct. The surface signals that reviewers spent a career calibrating on no longer correlate with defects.
What still works is the thing good reviewers always did and rarely wrote down: form your own model of what the change should look like before reading what it does look like, and then review the delta between the two. This playbook makes that explicit, adds the specific defect classes worth hunting, and — the part most reviews get wrong — separates what blocks from what doesn't.
When not to use this
Not for a two-line config change; just read it. Not as a substitute for CI — if you're manually checking formatting or type errors, fix the pipeline instead. And if the PR is over ~800 lines of non-generated code, don't review it: ask for it to be split. A thorough review of an 800-line PR isn't a thing that happens.
Did this work?
Every comment you left was either acted on or explicitly discussed, and no defect you could have caught reached production.
Tested on claude-opus-5. Evidence status is draft; it moves to battle-tested only on recorded runs, never by hand.