ワンクリックで
review-pr-comments
Review PR comments rigorously, present a justified summary, then post agreed responses individually
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 rigorously, present a justified summary, then post agreed responses individually
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
Create git commits with user approval, semantic commit format, and no Claude attribution
Implement technical plans from thoughts/shared/plans with verification
Create detailed implementation plans through interactive research and iteration
Test-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, or asks for test-first development.
Analyze and improve test code quality using Kent Beck's Test Desiderata framework. Use when analyzing test files, reviewing test code, identifying test quality issues, suggesting test improvements, or when asked to evaluate tests against best practices. Applies to unit tests, integration tests, and any automated test code.
Conduct multi-agent deep research on a topic with parallel web searches and synthesis
| name | review-pr-comments |
| description | Review PR comments rigorously, present a justified summary, then post agreed responses individually |
| model | opus |
| disable-model-invocation | true |
| allowed-tools | Bash(gh:*) Read Glob Grep |
| argument-hint | [PR number (optional — auto-detected from current branch)] |
Review all open PR comments, present a summary with your position on each, iterate with the user until agreed, then post responses individually to each comment.
PR target: if no PR number is given, auto-detect from the current branch. If none exists, ask.
Scope: only active comments (both inline and general). Ignore resolved and outdated ones.
Bias: defend the existing code. Only accept a change with a concrete technical reason (bug, violated principle, demonstrable readability gain). Reject style preferences without objective justification.
Context: before evaluating, read the full files affected plus related files (tests, types, direct dependencies).
Summary format: after reviewing every comment, present all decisions grouped by severity, one compact block per comment. Do NOT quote the full body verbatim — the user has GitHub open and can click through. Rephrase the claim in one sentence in your own words.
Extract severity from the comment's prefix when present ([SECURITY — CRITICAL], [CODE QUALITY — HIGH], [TESTS — LOW], [SUGGESTION], etc.). If no prefix, infer: correctness/security bugs → HIGH; performance, tests, or missing error paths → MEDIUM; style, docstrings, naming, "for symmetry", "for consistency" nits → LOW.
Block format per comment:
N. [ACCEPT|REJECT] {path}:{line or "general"} @{user} — {your one-sentence rephrase of the claim}
Reason: {1-2 lines, technical}
Fix: {concrete action you will take} # only if ACCEPT
Alternative: {what you propose instead} # only if REJECT
Link: {comment.html_url}
Present them in this order:
HIGH — decision required:
<blocks>
MEDIUM — decision required:
<blocks>
LOW / nits — proposed REJECT in batch (unless you rescue any):
<blocks, but with just the one-line rephrase + link; no Reason/Alternative>
Then ask ONE question: "Confirm my ACCEPT/REJECT on HIGH+MEDIUM and I discard the LOW batch — or do you want to rescue any LOW?"
Omit a section entirely if it has no comments. If ALL comments are LOW and every one would be REJECT, offer: "This round is all low-severity nits and rehashes of earlier positions. Close the round without replying?" — closing without a reply is a valid outcome when the original threads already carry your positions.
Exception: if a comment is genuinely short (<3 lines) and self-contained, quoting it verbatim is fine. The rule is "don't repeat information the user can read in one glance in GitHub".
Iteration: the user may push back on any decision. Update your positions and re-present the full summary until the user confirms agreement.
Publishing: once agreed, reply individually to each comment's thread (inline replies for inline comments, issue comment endpoint for general ones).
Language: responses posted to GitHub are always in English, regardless of the comment's language.