
pr-code-review
by EvolvingLMMs-Lab
pr-code-reviewは、機械学習とAI開発のためのスキルです。モデル構築から運用まで、包括的で効率的なAI開発ワークフローをサポートします。
ユースケース
AIモデル統合
LLMや機械学習モデルをアプリに統合。
プロンプト最適化
より良い結果を得るためのプロンプト改善。
データ分析自動化
AIを活用したデータ分析と洞察の抽出。
SKILL.md
name: pr-code-review description: Review GitHub PRs for bugs and code quality. Use when user says "review PR", "check PR #123", or mentions PR review.
PR Code Review
Straightforward code review for GitHub pull requests. Reviews the diff directly, only spawns agents when you need deeper codebase context.
Triggers
- "review PR #123"
- "check this PR"
- "code review"
How it works
You review the PR yourself using gh CLI. Only use explore agent if you need to understand unfamiliar parts of the codebase.
Review process
1. Check if PR is reviewable
gh pr view {number} --json state,isDraft,title,body
Skip if closed, draft, or already has your review:
gh api repos/{owner}/{repo}/pulls/{number}/comments | grep -c "Claude Code"
2. Get the diff and context
gh pr view {number}
gh pr diff {number}
Read CLAUDE.md if it exists - it contains project coding standards.
3. Review the changes
Look at the diff yourself. Focus on:
Real bugs - logic errors, missing error handling, runtime issues, incorrect variable usage
CLAUDE.md violations - if the project has one, check for type hints, docstrings, line length (88), PEP 8 naming, uv-only (no pip)
Skip these - pre-existing issues (lines not in diff), linter-catchable stuff, style nitpicks not in CLAUDE.md
4. Use explore agent only when needed
If you're unfamiliar with how the codebase does something, spawn an explore agent:
Task(subagent_type="explore", prompt="Find how error handling is done in lmms_eval/models/ - what patterns exist?")
Don't spawn agents for every PR. Most diffs are self-contained.
5. Post your review
Get the commit SHA for links:
gh pr view {number} --json headRefOid --jq '.headRefOid'
Post with:
gh pr comment {number} --body "..."
Comment format
Write like a helpful colleague, not a robot. Keep it brief and actionable.
When you find issues:
### Code Review
Took a look at this PR. Found a couple things worth addressing:
**Missing return statement** in `ovo_doc_to_target` - the function computes the value but never returns it.
https://github.com/{owner}/{repo}/blob/{SHA}/{file}#L102-L103
**No docstrings on public functions** - CLAUDE.md requires these for public APIs.
https://github.com/{owner}/{repo}/blob/{SHA}/{file}#L17-L45
Otherwise looks good. Let me know if you have questions.
When it looks good:
### Code Review
Reviewed the changes - looks good to me. Clean implementation, follows existing patterns.
When skipping:
Just tell the user why: "PR #123 is a draft, skipping review."
What NOT to flag
- Issues on lines you didn't modify (pre-existing)
- Formatting/linting issues (CI catches these)
- Personal style preferences not in CLAUDE.md
- Lines with
# noqaor similar ignore comments
Link format
Use full SHA, not short:
https://github.com/{owner}/{repo}/blob/{FULL_SHA}/{file_path}#L{start}-L{end}
Multiple PRs
> Review all open PRs
Get the list and review each:
gh pr list --state open --json number
スコア
総合スコア
リポジトリの品質指標に基づく評価
SKILL.mdファイルが含まれている
ライセンスが設定されている
100文字以上の説明がある
GitHub Stars 1000以上
3ヶ月以内に更新
10回以上フォークされている
オープンIssueが50未満
プログラミング言語が設定されている
1つ以上のタグが設定されている
レビュー
レビュー機能は近日公開予定です

