An agent skill for deep-dive review of a teammate's GitHub pull request — for cases that deserve more than a skim: infra/Terraform changes, IAM or access changes, or anything where "looks fine" isn't good enough. It traces claims against the actual repo (usages, history, other consumers of the changed code) instead of reasoning from the diff alone, cross-checks against existing PR discussion, triages findings down to the one issue that matters most, drafts and runs it through personify, and stages it as a GitHub pending review. It never posts, submits, or merges without explicit approval at each checkpoint.
It runs as a plain Markdown skill (SKILL.md), so any harness that supports
skill-style instructions can use it.
/plugin marketplace add smartwatermelon/pr-review
/plugin install pr-review@pr-review
Once installed, invoke it as /pr-review 123 (or /pr-review:pr-review 123
if you have another skill also named pr-review).
mkdir -p .claude/skills/pr-review
cp SKILL.md .claude/skills/pr-review/mkdir -p ~/.claude/skills/pr-review
cp SKILL.md ~/.claude/skills/pr-review/Reload or start a new session after installing.
The entire runtime artifact is SKILL.md. Any agent harness that loads
Markdown-based skills can use it by copying that one file into wherever the
harness expects skill definitions.
/pr-review 123
review PR 123
take a look at PR 456 in owner/repo
Not for your own working diff — use a plain code-review skill for that; this one is specifically for reviewing someone else's PR.
See SKILL.md for the full workflow: orient against the repo's own review
contract, classify the kind of change, verify claims against real usage
rather than trusting the diff, adversarially self-check candidate findings,
triage them down to the single highest-severity issue (rarely two), draft
and personify the review, then stage it as a pending GitHub review for a
human to submit.
pr-review is the first of three sibling skills that compose into one path from "review this PR" to a posted comment that reads like a person wrote it:
pr-review → personify → dumbify
(find it) (de-AI it) (compress it)
pr-review does the substance: it finds the issue, verifies it, and triages down to the one finding worth posting. The other two only touch how the draft reads.
- personify strips AI-writing
tells and, given a
VOICE.md, makes the comment sound like you specifically rather than like generically clean prose. - dumbify compresses the register further — terse, lowercase, fragment-heavy.
Phase 5 runs both automatically when they're installed, personify first, then dumbify. Both degrade gracefully, though not identically: if personify isn't installed, pr-review says so and shows the plain draft; if dumbify isn't installed, it skips that pass silently, because personify's work register already lands close to dumbify's default. Neither is load-bearing for the review's substance.
The two prose skills overlap, so running both is a choice rather than an upgrade. Personify's work register already produces lowercase starts and fragments; dumbify pushes past that. For most review comments personify alone is enough, which is why a missing dumbify costs you little. Order is fixed — personify's de-abstraction pass needs the full sentence that dumbify deletes.
MIT. See LICENSE.