ワンクリックで
review-pr-comments
Review PR comments, apply or reject each one with evidence, and post replies directly to GitHub
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Review PR comments, apply or reject each one with evidence, and post replies directly to GitHub
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Implement a GitHub issue step by step - ask about the branch, the PR target, and the phases first, then run the work and the review cycle with the review-pr skill. Use whenever the user asks to implement, work on, or execute a GitHub issue.
Create a well-structured GitHub issue with duplicate detection, code verification, and auto-labeling. Use whenever the user asks to create, file, open, log, or raise a GitHub issue, in any wording.
Review a GitHub pull request and post one clear review, with inline comments. The default is collaborative - you analyze deeply, explain the changes in simple terms, ask the user about each finding, and post a review that reflects their decisions. Use when the user asks to review a PR. --no-verdict = programmatic COMMENT-only mode that also writes .reviews/<PR_NUMBER>-review.md (used by implement-issue). --autonomous = unattended review that posts a formal verdict.
Root-cause a bug or a failed fix with evidence, not guesses - brainstorm causes, research them in parallel, confirm one with logs before changing code. Use when a reported behavior needs root-causing or a fix attempt has failed.
Rename or refactor a symbol safely by searching every naming-case variant across code, tests, docs, configs, and JSON. Use when renaming any identifier so no reference is missed.
Author or edit AI-facing instruction files in this repo - AGENTS.md, skills under .claude/skills/, subagents under .claude/agents/, and spawned-agent prompts. Covers the prompt-writing rules for current Claude models, the map-not-manual standard for AGENTS.md, the living-document rule, and when and where to persist a new rule. Use when editing any AI-instruction file.
| name | review-pr-comments |
| description | Review PR comments, apply or reject each one with evidence, and post replies directly to GitHub |
| argument-hint | ["pr-number-or-url"] |
Review PR (pull request) comments from reviewers. Make a careful decision on each one. Apply code changes where appropriate. Reply directly on GitHub.
If $ARGUMENTS contains a PR number or URL, use it. Otherwise, ask using AskUserQuestion:
"Which PR's comments do you want to review?"
Options:
Store as PR_NUMBER.
Gather the full context needed to evaluate comments:
# PR metadata
gh pr view $PR_NUMBER --json title,body,url,state,headRefName,baseRefName
# Full diff
gh pr diff $PR_NUMBER
# All review comments
gh api repos/$(gh repo view --json nameWithOwner -q .nameWithOwner)/pulls/$PR_NUMBER/comments --paginate
# Review-level comments (top-level review bodies)
gh api repos/$(gh repo view --json nameWithOwner -q .nameWithOwner)/pulls/$PR_NUMBER/reviews --paginate
If there are no comments, inform the user and stop.
Present a summary to the user:
For each PR comment, follow this structured decision framework:
Before deciding on a comment:
Search for evidence that supports or contradicts the comment:
| Area | Question |
|---|---|
| Architecture | Does the suggestion follow the project's existing patterns? |
| Types | Does it improve or maintain type safety? |
| Error handling | Does it improve resilience or correctness? |
| Security | Does it address a real vulnerability? |
| Tests | Does it improve coverage or correctness? |
| Code quality | Does it reduce duplication, improve naming, or simplify logic? |
| Scope | Is this within the PR's scope, or a broader refactor? |
For each comment, choose one:
Record 1-3 concise evidence bullets from the codebase to support your decision.
For each "Apply" or "Ambiguous -> Applied" decision:
cd web-projects/<project> && bun testtype: description): fix: address PR review - <short description>git push origin HEADAfter evaluating all comments and making changes, post a reply to each comment thread:
gh api repos/$(gh repo view --json nameWithOwner -q .nameWithOwner)/pulls/{pr_number}/comments/{comment_id}/replies -f body="<reply>"
Where {comment_id} is the id field (numeric databaseId) of the original comment.
Each reply must start with a bold decision tag:
**Applied** in <short-sha> - <one sentence describing what changed>.**Reject** - <1-3 sentences explaining why, with evidence>.**Ambiguous -> Applied** in <short-sha> - Interpreted as: <interpretation>. <what changed>.Keep replies short and factual.
-f body=.After posting all replies, present a summary to the user:
## PR Comments Review Summary
**PR:** #<number> - <title>
**Comments reviewed:** N
### Applied (N)
- Comment by @user on `file:line` - <what was changed>
### Rejected (N)
- Comment by @user on `file:line` - <why rejected>
### Ambiguous -> Applied (N)
- Comment by @user on `file:line` - Interpreted as: <interpretation>. <what was changed>
### Files Modified
- <list of files changed>
### Verification
- Tests: pass/fail/N/A