菜单
How to open pull requests in the DorkOS repo, including the automated-review controls (skip-review, review:light/deep, re-review). Use when opening a PR, deciding how much review a PR should get, or requesting a re-review after addressing feedback.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Develop, validate, and publish DorkOS marketplace packages — agents, plugins, skill-packs, and adapters. Use when creating marketplace items, working in the dork-labs/marketplace repo, or helping users build packages for the personal marketplace.
First-touch triage for incoming GitHub issues on dork-labs/dorkos. Use when DorkBot is asked to triage the issue queue: label each new issue by type and runtime, find likely duplicates, and ask the reporter for missing reproduction steps. Read-and-suggest by default; only comments or edits labels when explicitly allowed to act.
Regenerate DorkOS product screenshots and video loops from the real UI, and manage the shot registry, human overrides, and version archives that feed the marketing site, docs, and changelogs. Use when marketing/docs media is stale after UI changes, a new feature or docs page needs a money shot, a person wants to override an automated capture, or a release calls for a fresh capture or archive — anything under apps/site/public/product/ or the apps/e2e/capture pipeline.
Skill whose frontmatter name does not match its directory name.
Implements the Calm Tech design system using Tailwind CSS v4 and Shadcn UI. Use when writing styles, building components, or theming. For design decisions, see designing-frontend.
Writes human-friendly changelog entries and release notes. Use when populating changelog, preparing releases, or reviewing release notes quality.
| name | creating-pull-requests |
| description | How to open pull requests in the DorkOS repo, including the automated-review controls (skip-review, review:light/deep, re-review). Use when opening a PR, deciding how much review a PR should get, or requesting a re-review after addressing feedback. |
How DorkOS PRs are opened and how the automated Claude review behaves on them. This repo is routinely multi-agent, so the mechanics below keep PRs clean and the review loop cheap.
Code PRs come from an isolated worktree, never the shared main checkout (see the
working-in-worktrees skill). Base the worktree on origin/main, not local
main, so the PR diff contains only your changes:
git fetch origin
git gtr new <branch> --from origin/main --yes # the repo's worktree helper
Commit conventions and the pre-push gate live in the git:commit / git:push
commands. End commit messages with the Co-Authored-By trailer.
A PR with user-facing changes must include a changelog fragment under
changelog/unreleased/ — one file per change (<YYMMDD-HHMMSS>-<slug>.md; see
changelog/README.md). The post-commit hook usually writes one from your commit subject;
verify it exists and curate it (rewrite for a user, fix the category, add a (#PR) ref)
before opening the PR. Write one by hand if the hook skipped it or phrased it poorly. Do
NOT edit CHANGELOG.md's [Unreleased] section — it no longer holds entries; only
/system:release writes CHANGELOG.md.
Iterate as a draft, then mark ready when the branch is done:
gh pr create --draft --title "<type>(<scope>): <summary>" --body "<body>"
# ... push commits, iterate freely (no review runs while draft) ...
gh pr ready <number> # marking ready fires exactly one full review
Why draft-first: the auto-review is on-demand, not on every push (see below). A draft gets no review, so you can push freely; marking ready triggers one review of the final state. PR body: lead with what changed and why, link the spec or issue, and call out anything reviewers should look at first.
The claude-code-review workflow reviews on-demand, not on every push:
| Event | Review? |
|---|---|
| PR opened (non-draft) | One full review |
| Draft marked ready-for-review | One full review of the final state |
| New commits pushed | No auto-review (CI tests still run) |
re-review label applied | One re-review, scoped to the delta |
This mirrors how human teams work: pushes are work-in-progress, and the author pulls the reviewer back in with an explicit "ready again" signal. It avoids re-reviewing five or six times while you address feedback.
| Label | Effect |
|---|---|
skip-review | No automatic review at all. Apply at open (the workflow checks labels on the triggering event). |
review:light | Quick pass: only Important findings; skips nits and the deletion sweep. |
review:deep | Exhaustive: traces every caller, runs the full dangling-reference sweep. |
re-review | Request another pass after addressing feedback. Auto-cleared after the review runs, so re-apply it each time you want another look. |
Apply at creation so the opened event sees them:
gh pr create --draft --label review:light --title "..." --body "..."
Request a re-review after pushing fixes:
gh pr edit <number> --add-label re-review
# or, ad hoc, comment `@claude take another look` (handled by claude.yml)
Guidance: reach for skip-review only on genuinely trivial PRs (typo, version
bump) where you are the merger and have full context. Prefer review:light over
skip-review when in doubt: you still get Important-only coverage at low cost.
Use review:deep for risky changes (security, migrations, broad refactors,
deletions).
The four labels must exist in the repo before they can be applied. Create them once:
gh label create skip-review --description "Skip the automated Claude review" --color ededed
gh label create review:light --description "Quick review: Important findings only" --color fbca04
gh label create review:deep --description "Exhaustive review" --color b60205
gh label create re-review --description "Request another automated review pass" --color 0e8a16
.github/workflows/claude-code-review.yml or REVIEW.md only take effect after
merge. Merge to main, then exercise on a throwaway PR.post-commit hook writes a changelog fragment under
changelog/unreleased/ from the commit subject (it dedupes across amend/rebase and
never touches CHANGELOG.md). For changes that should not land in the user-facing
changelog, touch .claude/.changelog-populator.lock before committing (the lock is
gitignored) and delete any fragment it already wrote.