A spec driven build pipeline for Claude Code. Every feature goes from a brief to a spec with acceptance criteria, then through attention to detail, QA and handoff gates, so nothing ships half done or unpolished.
claude plugins marketplace add mindantic/build-skill
claude plugins install build@build-local| Command | What it does | Writes |
|---|---|---|
/build:build |
Works out the next step and runs it | |
/build:brief |
Collects the brief | docs/brief.md, docs/project.md, docs/decisions.md |
/build:research |
Competitor analysis | docs/research/<competitor>.md |
/build:workflow |
How tasks run, project rules | docs/workflow.md, docs/rules.md |
/build:design |
Design direction from inspiration | docs/design.md |
/build:dev |
Technical decisions | docs/dev.md |
/build:srs |
Phases, features and tasks with acceptance criteria | docs/specs/phase-NN-<slug>.md |
/build:execute |
Builds a phase or task | code |
/build:atd |
Attention to detail gate | docs/evidence/task-NNN/atd.md |
/build:qa |
Functional QA gate | docs/evidence/task-NNN/qa.md |
/build:handoff |
Final gate, marks tasks Done | |
/build:check |
Current status and next step |
Every command works with no arguments. Claude asks one question at a time, each with options, a recommendation, and why or why not, until nothing is left to guess.
- Only
/build:handoffcan mark a task Done, and only when atd and qa both passed on the current code. - Every expected behavior is an acceptance criterion. If it is not one, it is not built or checked.
- Gates verify in a real browser with agent-browser at 375, 768, 1024 and 1440, and every pass needs evidence.
reference/baseline.mdis the global detail standard. A project'sdocs/rules.mdextends it.
MIT