| name | review |
| description | PR review skill — fetches active GitHub Copilot review comments on a PR, critically evaluates each suggestion, implements those that improve code posture, and dismisses those that don't add value. |
| model | opus |
| disable-model-invocation | true |
/review
Processes active review comments on a pull request. Every suggestion is challenged
for relevance and value. Worthwhile suggestions are implemented; weak ones are
dismissed with a clear rationale.
Supporting scripts
| Script | Purpose |
|---|
${CLAUDE_SKILL_DIR}/scripts/fetch-threads.sh <owner> <repo> <number> | Paginated GraphQL — returns unresolved review threads |
${CLAUDE_SKILL_DIR}/scripts/resolve-thread.sh <thread-node-id> | GraphQL mutation — marks a thread resolved |
${CLAUDE_SKILL_DIR}/scripts/react-comment.sh <owner> <repo> <comment-id> <+1|-1> | Posts a thumbs-up or thumbs-down reaction |
${CLAUDE_SKILL_DIR}/scripts/reply-comment.sh <owner> <repo> <number> <comment-id> <body> | Posts an inline reply to a thread |
Input
$ARGUMENTS — a PR number (e.g. 42) or a GitHub PR URL.
Steps
1. Resolve PR number
Extract the PR number from $ARGUMENTS. If a URL is provided, parse the number
from it.
2. Fetch PR metadata
gh pr view <number> --json number,title,headRefName,baseRefName,state,url,owner,repository
Confirm the PR is open. If it is merged or closed, stop and tell the user.
Extract owner and repo name from the PR metadata (or from gh repo view --json owner,name)
for use in subsequent API calls.
Checkout the PR branch so edits can be committed:
gh pr checkout <number>
3. Fetch unresolved review threads
${CLAUDE_SKILL_DIR}/scripts/fetch-threads.sh <owner> <repo> <number>
Filter to threads where the first comment's author.login is "copilot-pull-request-reviewer"
— the script already excludes resolved threads.
Each thread node contains its id (for resolution), path, line, and all
comment bodies. No separate grouping step is needed.
If there are no matching threads, print "No active Copilot review comments found."
and stop.
4. Evaluate and action each thread — one at a time
For every thread, perform the following evaluation loop before moving to the next.
4a. Read the affected file and context
Use the first comment's databaseId from the GraphQL response as <comment-id> for
all REST calls (reactions, replies) in this thread.
gh api repos/{owner}/{repo}/pulls/comments/<comment-id> \
--jq '{path, line, diff_hunk}'
Read the file at the commented path to understand the surrounding code.
4b. Critically evaluate the suggestion
Ask the following questions:
- Is it correct? Does the suggestion fix a real bug or address a genuine issue,
or is it based on a misreading of the code?
- Does it improve code posture? Security, correctness, readability, performance,
or maintainability — does the change make the code measurably better?
- Is it in scope? Does it touch what the PR is actually changing, or is it a
general cleanup unrelated to the PR's intent?
- Does it align with project conventions? Check CLAUDE.md — if the suggestion
contradicts established patterns (Biome, Tailwind v4, TanStack patterns), it is
likely wrong.
- Is it non-trivial noise? Purely stylistic opinions, unnecessary complexity,
premature abstractions, or suggestions that duplicate what the linter already
enforces are noise.
A suggestion must pass at least questions 1 and 2 to be implemented.
4c. Implement or dismiss
If the suggestion is worth implementing:
- Apply the change to the affected file.
- React 👍 to the Copilot comment to signal acceptance and provide Copilot fine-tuning
feedback:
${CLAUDE_SKILL_DIR}/scripts/react-comment.sh <owner> <repo> <comment-id> +1
- Resolve the thread using the thread node ID obtained in step 3:
${CLAUDE_SKILL_DIR}/scripts/resolve-thread.sh <thread-node-id>
If the suggestion should be dismissed:
- Do NOT touch the code.
- React 👎 to the Copilot comment to signal rejection and provide Copilot fine-tuning
feedback:
${CLAUDE_SKILL_DIR}/scripts/react-comment.sh <owner> <repo> <comment-id> -1
- Reply to the comment thread with a concise, direct rationale:
${CLAUDE_SKILL_DIR}/scripts/reply-comment.sh <owner> <repo> <number> <comment-id> \
"Closing — <reason: e.g. 'the existing code is correct because...', 'this contradicts the project'\''s Biome config', 'premature abstraction not warranted here'>."
${CLAUDE_SKILL_DIR}/scripts/resolve-thread.sh <thread-node-id>
5. Commit and push (if changes were made)
git add -p
git commit -m "fix: address PR review comments (#<number>)"
git push
6. Output summary
Print a table:
| File | Line | Decision | Reason |
|---|
src/... | 42 | Implemented | Fixed null-safety issue in ... |
src/... | 78 | Dismissed | Stylistic preference; Biome handles formatting |
Then print the count: X implemented, Y dismissed.
Constraints
- Challenge every suggestion — do not rubber-stamp Copilot output.
- Never implement a suggestion that contradicts project conventions (CLAUDE.md).
- Never implement a suggestion that introduces complexity without a clear benefit.
- If a suggestion is ambiguous, lean toward dismissal and explain why in the reply.
- Do not modify files outside the PR's changed file set unless a suggestion is
clearly a bug fix that warrants it.
- Always respond to every comment thread — react and resolve for implemented
suggestions; reply with a rationale and resolve for dismissed ones.
