Pr review
Review someone else's pull request: correctness, scope, tests, and an actionable verdict.From its SKILL.md
npx -y skills add ayaqen/Agents-Scripts --skill pr-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.0 KB, as published. Nobody here has run it
pr-review
Reviewing code you didn't write — including AI-generated PRs — with one job: protect the main branch while keeping the contribution, and the contributor, moving. Covers someone else's diff; for your own, see self-review.
When to use
- Asked to review a pull request.
- Merging an external contribution.
- Checking agent-generated work before it lands.
Workflow
- Context before code: read the PR description, the linked issue, and CI status first, so you know what the change claims to do before judging whether it does it.
- Read the diff hunk by hunk for correctness — boundaries, error paths, concurrency. Then read the negative space: missing tests, un-updated callers, docs/changelog for user-visible changes.
- When stakes warrant, verify claims instead of trusting them: check out the branch, run the tests, exercise the feature yourself.
- Treat generated code with extra suspicion — confident-but-wrong logic, invented APIs, tests that assert nothing. Confirm new tests actually fail if the change is reverted.
- Classify every finding explicitly as blocking (correctness, security, data loss) or non-blocking (style, preference). A review without this split stalls the PR on nits.
- Deliver an explicit verdict with rationale. Keep comments line-anchored, respectful, and paired with a concrete fix. For small fixes, prefer fixing and merging yourself over another review round-trip.
- On merge, preserve contributor credit with a
Co-authored-bytrailer.
Pitfalls
- Approving because CI is green — CI proves what it tests, not whether the design is right.
- Drowning a real design flaw under a pile of style nits.
- "Needs work" without saying what work — every blocking comment names the fix.
- Reviewing the author instead of the code.
- Skimming a huge PR end to end. A review you can't actually perform protects nothing — request a split instead.