[REPLACE: SKILL_NAME] Skill
对 [REPLACE: WATCHED_REPO] 新开的 PR 做首轮审查——给出结论 + 友好的评论 + 打标签
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
${var} — Optional. PR number to review. If empty, scans all newly-opened PRs on [REPLACE: WATCHED_REPO].
Today is ${today}. Review external PRs on [REPLACE: WATCHED_REPO] with a focus on [REPLACE: REVIEW_FOCUS].
Steps
- List candidates — every open PR opened in the last 24h that hasn't been reviewed by this skill yet:
If ${var} is set, that PR number is the only candidate. Otherwise get the cutoff with its own call, date -u -d '24 hours ago' +%Y-%m-%dT%H:%M:%SZ, and write the printed value into the filter literally (the below):
gh pr list -R [REPLACE: WATCHED_REPO] --state open --json number,author,createdAt,additions,deletions --jq '.[] | select(.createdAt >= "<SINCE>") | select(.author.login != "github-actions[bot]" and .author.login != "aeonframework") | .number'
Track previously-reviewed PRs in memory/topics/[REPLACE: SKILL_NAME]-reviewed.json (a flat array of PR numbers). Skip anything already in there.
- For each PR — fetch metadata + diff:
One call per PR, with its number written in literally (the below). Read the output directly; do not save it to a file:
gh pr view <PR> -R [REPLACE: WATCHED_REPO] --json title,body,additions,deletions,files,author
gh pr diff <PR> -R [REPLACE: WATCHED_REPO]
Run the diff call only for a PR under the size limit. Skip if additions + deletions > [REPLACE: MAX_PR_LINES]: flag it as DEFERRED: too large for first-touch review, leave a note, and move on.
- Apply the rubric — assign one of four verdicts:
| Verdict | Trigger |
|---------|---------|
| ACCEPT | Touches expected paths, follows repo conventions, focused scope, no obvious bugs in the diff. |
| NEEDS-CHANGES | Reasonable intent but specific issues: missing tests, broken format, incorrect assumption, naming. |
| DEFER | Out of scope for this skill — needs a human reviewer (large refactor, architectural change). |
| OUT-OF-SCOPE | Touches files outside what the repo accepts contributions on (e.g. lock files, generated assets). |
Focus the rubric on [REPLACE: REVIEW_FOCUS] — that's the lens that matters most for this repo.
- Post a comment — use a friendly, specific tone. Acknowledge the contributor, name the verdict, give 1-3 concrete bullets:
Write the comment to /tmp/review-comment-.md with Write:
Thanks for the PR! [verdict text]
- [bullet 1]
- [bullet 2]
Then post it:
gh pr comment <PR> -R [REPLACE: WATCHED_REPO] --body-file /tmp/review-comment-<PR>.md
- Label the PR via
gh pr edit -R [REPLACE: WATCHED_REPO] --add-label ": useaccepted/needs-changes/defer/out-of-scope(create the labels in the target repo first if they don't exist).
- Notify via
./notifyonly onACCEPTorOUT-OF-SCOPE— those are the actionable verdicts for the operator. Silent onNEEDS-CHANGESandDEFER(the comment on the PR is the signal).
- Log — append to
memory/logs/${today}.md:
## [REPLACE: SKILL_NAME]
- **PRs reviewed**: N (skipped M as previously seen)
- **Verdicts**: accept=X, needs-changes=Y, defer=Z, out-of-scope=W
- **Status**: REVIEW_OK | REVIEW_QUIET (no new PRs)
Network note
gh handles auth via the workflow's GITHUB_TOKEN. To comment on or label a PR in [REPLACE: WATCHED_REPO], the token needs pull-requests: write and issues: write permission on that repo — verify the workflow grants those, or this skill will silently fail to write.
Shell shape. Each Bash call is one allowed command (pipes are fine). The run refuses a call that sets a variable (PRS=$(gh pr list ...)), uses if / for, joins commands with ;, && or ||, or redirects into a file (> .pr-meta.json), and each refusal costs a model turn. So make one gh call per PR with the number written in literally, read gh output directly, and write the comment body, reviewed.json and the log with Write/Edit.
Constraints
- Be welcoming. The PR may be someone's first open-source contribution. Lead with thanks, then specifics.
- Never auto-merge. This skill is first-touch review, not auto-merge. Even an
ACCEPTverdict still waits for a human merge. - Idempotent. Re-running the skill must never double-comment. The
reviewed.jsonstate file is what enforces that.