메뉴
Surface open PRs that have genuinely unaddressed review feedback, filtering out CodeRabbit auto-summaries and CI-bot noise. Use before ending a session, when asked "which of my PRs need attention", or to gate a session close on unresolved review threads.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Sync git repository with upstream. Use at the start of a session, when asked to sync, get up to date, check git status, or when working on a stale branch.
Analyze content and generate illustrations via Gemini image API
Brainstorm multiple visual directions for a blog image, generate them in parallel, build a comparison page, and optionally publish as a shareable link (Surge.sh or gist).
Bulk-parallel CLIs — turn N sequential gh/bd/git/file tool calls into a single fan-out JSON call. Use when the session is about to fire ≥3 similar sequential calls (gh pr view, bd show, Read of small files, up-to-date diagnose across repos).
Extract durable lessons from a completed Claude session and codify them in the right CLAUDE.md files or skills. Use at the end of a long session, after a bug hunt that surfaced a non-obvious constraint, or when the user asks "what can we learn from this session". Discovers CLAUDE.md files dynamically, routes lessons by generic scope (project / shared conventions / environment / machine-local), enforces neutral voice, and asks for approval before editing.
Read (not generate) images via a subagent so the pixel payload stays out of the parent conversation. Dispatches Haiku by default, escalates to Sonnet/Opus when the caller flags ambiguity or needs load-bearing detail.
| name | pr-hygiene |
| description | Surface open PRs that have genuinely unaddressed review feedback, filtering out CodeRabbit auto-summaries and CI-bot noise. Use before ending a session, when asked "which of my PRs need attention", or to gate a session close on unresolved review threads. |
| allowed-tools | Bash, Read |
git-safe-push guards one PR footgun (pushing to a merged branch). This guards
its sibling: open PRs quietly piling up with review comments you never
resolved or replied to. GitHub gives no nudge, and a naive "the last comment
isn't mine" filter over-flags ~3× — it can't tell a CodeRabbit auto-summary or a
CI-bot ping from a real ask.
pr-hygiene enumerates your open PRs across GitHub identities and sorts them
into three tiers so you act only on the ones that need it.
It is a single self-bootstrapping uv run --script file — no packaging needed.
Symlink it onto $PATH the same way as git-safe-push:
ln -sf ~/gits/chop-conventions/skills/pr-hygiene/pr_hygiene.py ~/.local/bin/pr-hygiene
The PEP 723 shebang resolves typer on first run. Requires gh (authenticated).
pr-hygiene # ranked markdown table (🔴 first)
pr-hygiene --json # machine-readable JSON (+ counts, errors)
pr-hygiene --repo idvorkin/chop-conventions # filter to one repo
pr-hygiene --author octocat # override author (repeatable)
pr-hygiene --no-fail # always exit 0 (don't gate on 🔴)
Default authors are idvorkin and idvorkin-ai-tools. Bitbucket repos (e.g.
igor2) are invisible to gh and simply aren't listed. Per-PR GraphQL failures
are reported inline (tier error), never crash the run.
For each PR one GraphQL round-trip pulls review threads (isResolved),
reviewDecision, reviews, and issue comments. Tiers:
reviewDecision == CHANGES_REQUESTED, OR a human reviewer has the last word
and the author hasn't responded since (push, comment, or review).Two noise rules make this work:
Cross-identity nuance: a comment counts as a reviewer ask only when its author
differs from the PR author — so a self-note on your own PR is ignored, while a
thread left by your other identity (e.g. idvorkin reviewing an
idvorkin-ai-tools PR) correctly counts.
The classification logic lives in pure, unit-tested functions in
pr_hygiene.py (see tests/); gh shell-outs are injected so tests run
offline.