Agent Skills

Since v1.12, npx buildcrew also installs 7 Agent Skills (opens in a new tab) into .claude/skills/ — the same pipelines packaged in the open SKILL.md standard supported by Claude Code, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI and 40+ other tools.

💡

One install, every tool. Role prompts stay in .claude/agents/*.md as the single source of truth — skills carry the pipeline order and discipline rules, then dispatch subagents where supported (Claude Code) or adopt each role sequentially in single-context tools.

The 7 Skills

Skill	Covers	Modes wrapped
`buildcrew`	Full feature pipeline — plan → challenge → design → challenge → dev → [QA ∥ browser ∥ spec-verify] parallel fleet → review → coherence	Feature
`buildcrew-qa`	Browser QA via Playwright, 3-lens QA audit on diffs, 0-10 health score	Browser QA, QA Audit, Health
`buildcrew-security`	OWASP Top 10 + STRIDE audit with a fix loop for critical findings	Security
`buildcrew-debug`	4-phase root cause investigation, verified by QA	Debug
`buildcrew-review`	Code review with auto-fix, pre-code architecture review, UX review	Review, Arch Review, Design Review
`buildcrew-ship`	Pre-flight tests → version bump → changelog → PR	Ship
`buildcrew-think`	6 forcing questions ending in a design doc	Think

Versioned sync

Skills use the same version-aware update logic as agents:

New install → copied to .claude/skills/<name>/SKILL.md
Version bump upstream → updated in place
Local edits → preserved until the skill actually changes upstream
npx buildcrew --uninstall removes only package-shipped skills; your own skills are untouched

Governance

buildcrew treats agents like production actors:

Least privilege — every agent declares its own tools: frontmatter. QA and audit roles are read-only; only developer writes code; only shipper touches git push.
Deny by default for the dangerous stuff — the permission preset (--install --with-permissions) denies rm -rf, sudo, force-push, git reset --hard, git clean -fd, and npm publish/npm unpublish. Registry publishing always requires explicit human approval.
Audit trail — every dispatch, completion, and file write lands in .claude/buildcrew/events.jsonl. Review live with npx buildcrew watch, or after the fact with npx buildcrew report. Handoff Records plus the coherence score prove coordination at the artifact level.
Model governance — model: aliases (opus, sonnet) resolve to the current model generation automatically. Pin or downgrade a role by editing its installed agent file.

thinker Overview