Ответы на комментарии ревью
Разбирает комментарии ревьюера к PR: проверяет каждое утверждение по коду, исправляет то, что верно, возражает с доказательствами там, где неверно, и готовит ответ на каждый комментарий. Используйте, когда пользователь прислал комментарии ревьюера, которые нужно отработать.
review-feedback-responseУстановка
Выполните одну команду — она сохранит SKILL.md в папку скиллов агента. Перезапустите агента, и он подхватит скилл сам.
mkdir -p ~/.claude/skills/review-feedback-response && curl -fsSL https://anymodel.org/en/skills/code-review/review-feedback-response/SKILL.md -o ~/.claude/skills/review-feedback-response/SKILL.md
mkdir -p ~/.agents/skills/review-feedback-response && curl -fsSL https://anymodel.org/en/skills/code-review/review-feedback-response/SKILL.md -o ~/.agents/skills/review-feedback-response/SKILL.md
New-Item -ItemType Directory -Force "$HOME\.claude\skills\review-feedback-response" | Out-Null; Invoke-WebRequest https://anymodel.org/en/skills/code-review/review-feedback-response/SKILL.md -UseBasicParsing -OutFile "$HOME\.claude\skills\review-feedback-response\SKILL.md"
New-Item -ItemType Directory -Force "$HOME\.agents\skills\review-feedback-response" | Out-Null; Invoke-WebRequest https://anymodel.org/en/skills/code-review/review-feedback-response/SKILL.md -UseBasicParsing -OutFile "$HOME\.agents\skills\review-feedback-response\SKILL.md"
Другой агент: сохраните файл в его папку скиллов или дайте ему ссылку на файл: https://anymodel.org/en/skills/code-review/review-feedback-response/SKILL.md
SKILL.md
Review Feedback Response
Treat each reviewer comment as a claim to check, not an order to obey and not an attack to defend against. Fix what is true, explain what is not, and keep the thread easy to resolve.
1. Collect all comments
Get every open comment with its file, line and thread context.
gh pr view 123 --comments
gh api repos/OWNER/REPO/pulls/123/comments --paginate # inline review comments
git log --oneline origin/main..HEAD
If the user pasted the comments, work from those. Number them (1, 2, 3 ...) and keep that numbering in every later step.
2. Verify each claim
For each comment, before deciding anything:
- Open the code the comment points at, plus the code around it and its callers.
- Restate the reviewer's claim in one sentence.
- Try to confirm it: reproduce the bug, write a failing test, trace the input, read the docs of the API in question.
- Classify it:
- correct: the code is wrong or unclear as stated.
- partly correct: real concern, but the proposed fix is not the best one.
- incorrect: the reviewer misread the code or is missing context.
- preference: a matter of taste, not correctness.
- question: needs an answer, not a change.
Never agree with a claim you have not checked, and never dismiss one without reading the code.
3. Act on each comment
- Correct: fix it with the smallest change; add or adjust a test that fails before the fix when it is a behavior bug.
- Partly correct: fix the real concern your way and explain why.
- Incorrect: do not change the code. Reply with evidence (a code reference, a test result, a doc quote, a command output).
- Preference: follow the repo convention if one exists; otherwise ask the user whether to accept it. Cheap and harmless suggestions are usually worth accepting.
- Question: answer it; if the answer reveals the code is confusing, add a clarifying comment or rename.
Ask the user before pushing back on a senior reviewer, on a blocking comment, or when the decision has product or architecture consequences.
4. Verify the whole thing again
After the edits, run the project's formatter, linter, type check and the test suite for the affected area. Confirm that fixing one comment did not break another. Keep fixes in a new commit (or fixup commits) so the reviewer can see exactly what changed. Do not rewrite history of a shared branch without asking.
5. Draft replies
One reply per comment, short, in the same order. Use this shape:
#1 (src/api.py:40) - fixed in <short hash>. Added the missing null check and a test for an empty payload.
#2 (src/api.py:77) - kept as is. The retry is safe here because the call is idempotent (see docs/api.md, "Retries"); a test covers a double delivery.
#3 (README.md) - done, thanks.
#4 (src/util.py:12) - question: do you want this in this PR or a follow-up? I would do a follow-up to keep the diff small.
Rules for replies:
- Lead with the outcome (fixed / not changed / need input), then the reason.
- Reference commits, tests or docs; do not argue from authority or feelings.
- Be polite and specific; thank only when it is genuine, and skip filler.
- Never mark a thread resolved on the reviewer's behalf unless the team convention says the author does it.
6. Final report to the user
Addressed: N of M comments
Fixed: #1, #3
Disagreed (with evidence): #2
Need your decision: #4
Tests run: <command and result>
Commits: <hashes>
Pitfalls
- Do not blindly apply every suggestion; some contradict each other or the codebase.
- Do not defensively reject comments to save work.
- Do not expand the PR with unrelated fixes while addressing feedback; open a follow-up instead.
- Do not force-push, resolve threads, or re-request review without the user's go-ahead.
- Do not post replies yourself unless the user asked; drafts first.
Source: AnyModel skills catalog · License: CC BY 4.0
Подключить через AnyModel API
Скилл работает на той модели, которую использует агент. Подключите Claude Code, Codex или Cursor к AnyModel и запускайте его на Claude, GPT, Gemini и других моделях по одному ключу — оплата за токены, без подписки.
AnyModel