Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/adamayoung/TMDb --skill pr명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Implement features and fix bugs test-first, the Canon TDD way (test list → one test → make it pass → refactor → repeat). Use when implementing any new feature, endpoint, model, or method, fixing a bug, or executing a plan — write the test list and a failing test BEFORE production code.
Capture durable, project-specific learnings from the work just done into the knowledge/ base — gotchas, implementation quirks, things looked up or web-searched, live-API behaviours, and design decisions (as ADRs). Use before opening a PR (it runs automatically in /deliver), or any time you've learned something worth remembering. Records only non-obvious, reusable facts; skips anything already in the repo, CLAUDE.md, or git history.
Take a plan all the way to a ready-to-merge pull request — review the plan (scaled to risk), implement it test-first, code-review and fix, run the CI gate, open the PR, and watch it green. Use after you have an approved plan (from plan mode, or the Plan agent); pass `issue <n>` to deliver a specific GitHub issue you name, or `next` to take the top startable issue off the project board's Ready column — either way it re-verifies, claims and plans the issue first. Invoking it with a plan already in hand is itself plan approval — it then runs autonomously to a single hard stop: ready-to-merge.
SKILL.md 표시 중
| name | pr |
| description | Create a pull request |
I'll create a pull request for the current branch by following these steps. If any steps fail, stop.
Mode — check the arguments passed to this skill (shown at the end). If they
include reviewed (or skip-review), an upstream step already ran the review
and fixed its findings (e.g. /deliver's code-review phase via
/review-changes), so skip steps 5–7 to avoid reviewing identical code
twice. Otherwise (standalone /pr), run the internal review as normal.
Also skip steps 5–7 when the branch changes no reviewable code. Use
/review-changes' own gate set, not a narrower one — committed scripts under
.claude/workflows/ and Scripts/ are reviewable too, and a PR editing only
Scripts/check-defaulted-witnesses.py changes a CI gate:
git diff --name-only origin/main...HEAD \
| grep -qE '\.swift$|^\.claude/workflows/|^Scripts/' \
|| echo "no reviewable code → skip review"
(If the change rewrites the pipeline's own skills, don't skip — pass
force-review to /review-changes; see its §0.)
Run /format to format code
Commit all outstanding work. Run git status. If the working tree has any uncommitted or unstaged changes (the feature work, plus the formatting from step 1), stage and commit them — the PR reflects committed history only, so anything left uncommitted will be missing from the PR. First verify no secrets, .env, or build artifacts are included (per CLAUDE.md — .env must stay gitignored; check git status before git add). Use a descriptive gitmoji message (or several logical commits if the work spans concerns); formatting-only changes can use "🎨 apply code formatting". If the tree is already clean (e.g. work was committed during /deliver), this is a no-op.
Rebase onto the latest origin/main. Fetch the remote and rebase the feature branch directly onto origin/main — do this before make ci so the gate (and the eventual PR) reflects the real merge result, not stale code:
git fetch origin
git rebase origin/main
origin/main, not local main — this is worktree-safe. A /deliver runs inside a git worktree, and git checkout main there fails (fatal: 'main' is already used by worktree …) whenever the main checkout is on main; rebasing onto origin/main avoids checking main out at all and uses the true remote base directly.git rebase reports conflicts, stop, resolve them, then continue. Never skip or force past a conflict you don't understand.git push --force-with-lease at the push step below.origin/main with nothing to replay → this is a fast no-op.Run the pre-PR gate — MANDATORY; it must pass before going further. Run it directly (do not delegate). If it fails, stop and fix — commit the fixes — then re-run; never open a PR on a red gate. Scale the gate to the diff:
Default — full make ci. The gate CLAUDE.md requires before any PR: lint, markdown lint, unit tests, integration tests, the release build, and the docs build. Use this whenever any code or build-affecting file changed.
Docs/config-only fast gate. When the diff touches no build- or test-affecting files — i.e. no *.swift and none of Makefile, Package.swift, Package.resolved, *.xctestplan, .github/workflows/**, or Tests/TMDbTests/Resources/** — the test/release-build legs of make ci have nothing to exercise. Run only the meaningful checks instead: make lint && make lint-markdown && make build-docs (drop build-docs if no *.docc/** changed). Detect this with:
git diff --name-only origin/main...HEAD \
| grep -qE '\.swift$|^Makefile$|^Package\.(swift|resolved)$|\.xctestplan$|^\.github/workflows/|^Tests/TMDbTests/Resources/' \
&& echo "code/build touched → full make ci" \
|| echo "docs/config-only → fast gate"
JSON fixtures count as test-affecting. They are a build input of the TMDbTests target via .process("Resources") (Package.swift), so a fixture-only diff still has to run the unit suites — a re-captured value can break a decode assertion that no amount of lint will catch. .github/workflows/ci.yml's swift paths filter carries the same path for the same reason; keep the two in step.
The PR's own CI still runs the full matrix regardless, so this only trims the local gate; it never lowers what actually guards main. When in doubt, run full make ci.
Re-lint new Swift files without the cache. make ci's lint leg is (Makefile), and SwiftLint (). On files this has been seen to report a locally — passing / that the PR's CI job (a clean checkout, no cache) then fails on. So (), run a cache-bypassing lint before trusting the gate:
(skip in reviewed mode or when no reviewable code changed) Run the /review-changes skill — it scales the review to the diff (a single code-reviewer for a small change, the fan-out + adversarial-verify Workflow for a large one), follows .github/CODE_REVIEW.md, and returns a severity-graded report
(skip in reviewed mode or when no reviewable code changed) Summarize the code review findings:
(skip in reviewed mode or when no reviewable code changed) If there are critical/high-severity issues:
Ensure a clean working tree before pushing. Re-run git status; commit anything still outstanding (e.g. review fixes from steps 5–7, or make ci fixes) so the push includes everything. The tree must be clean before proceeding. Then run git diff origin/main...HEAD to understand all changes going into the PR.
Analyze the commits and changes to generate an appropriate title and summary
Push the current branch to remote (git push -u origin <branch> if not yet pushed; otherwise git push — use git push --force-with-lease if you rebased in step 3)
Create a PR with the GitHub MCP — mcp__github__create_pull_request (owner/repo from the origin remote, base: main, head: <branch>, plus title/body). The branch must already be pushed (step 10). If the call fails with 401/403 (PAT expired or missing scope), fall back to gh pr create with the same title/body.
Title — MUST start with a gitmoji prefix matching the change:
| Emoji | Code | Use for |
|---|---|---|
| ✨ | :sparkles: | New features |
| 🐛 | :bug: | Bug fixes |
| 📝 | :memo: | Documentation updates |
| ♻️ | :recycle: | Refactoring |
| ✅ | :white_check_mark: | Adding/updating tests |
| 🔧 | :wrench: | Configuration changes |
| ⚡️ | :zap: | Performance improvements |
| 🎨 | :art: | Code style/formatting |
Example: ✨ Add createdBy property to TVSeries.
Body — fill this skeleton, keeping only the sections that apply, and end with the attribution line exactly as shown.
Closes #NNN is required whenever the work has a tracked issue — drop the line only when it genuinely has none. It is not decoration: GitHub uses it to close the issue on merge, which is what carries the board item to Done, and /watch-pr reads it to move that item to In review when the PR goes green (see .github/ISSUE_FILING.md → Board status — the column lifecycle). Omit it and the issue stays open after merge and never leaves its column — silently, since nothing fails.
Closes #NNN.
[Brief description of what this PR does and why]
✨ [Description]
📝 [Description]
✅ [Description of test coverage]
📚 [Description]
: [Description]
🤖 Generated with []()
$ARGUMENTS
swiftlint --strict .~/Library/Caches/SwiftLint.swiftfile_lengthtype_body_lengthLint.swift filegit diff --name-only --diff-filter=A origin/main...HEAD | grep -q '\.swift$'swiftlint lint --strict --no-cache .
Fix anything it flags (oversized files: split, or add a // swiftlint:disable directive matching the AccountService+Pagination.swift precedent) and re-run. This catches file-size violations locally instead of on a CI round-trip.
The harness appends a claude.ai/code session link after the attribution
line — that is expected; leave it in place.