Skip to content

Latest commit

 

History

15 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

pr-review

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.

Installation

Claude Code plugin

/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).

Claude Code, project-local

mkdir -p .claude/skills/pr-review
cp SKILL.md .claude/skills/pr-review/

Claude Code, global

mkdir -p ~/.claude/skills/pr-review
cp SKILL.md ~/.claude/skills/pr-review/

Reload or start a new session after installing.

Any other harness

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.

Usage

/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.

What it does

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.

Works with personify and dumbify

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.

License

MIT. See LICENSE.

About

A Claude Code skill for deep-dive review of a teammate's GitHub pull request

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages