A Claude Code skill that delivers a feature the way the agent-center project did: design once, then let a team of agents build and check each other with no human in the loop.
Two phases, two drivers:
| Phase | Driver | Output |
|---|---|---|
| 1. Design (human-led, ~1–2h) | You + Claude, conversational | DESIGN.md — an unambiguous decision ledger (ADRs + task DAG + acceptance gates) so agents never need to ask a human mid-build |
| 2. Build (hands-off) | A deterministic Workflow script | Per task: own git worktree → Dev opens a PR → independent Tester verifies the gate → Supervisor §-1 review → a single Integrator merges to trunk. Failures rework on the same branch; persistent failures mark needs-human without blocking other branches. |
The core principle, proven by the source project: the agent that writes code is never the agent that approves it. Separation of duties is enforced by the workflow's structure — physical isolation via git worktrees, PRs, and a single merge authority — not by hoping an agent behaves.
SKILL.md— the skill: how Claude runs the design session and launches the build.feature-factory.workflow.js— the hands-off build pipeline (worktree → Dev/PR → Tester → §-1 review → Integrator-only merge).feature-factory.workflow.test.mjs— E2E orchestration tests (run withnode feature-factory.workflow.test.mjs).
Copy this directory into ~/.claude/skills/feature-factory/, then invoke with /feature-factory.
The orchestration logic is covered by 15 passing E2E tests (mocked agents, real script). Not yet validated against a real git repo + live agents end-to-end — that's the next step.
MIT — see LICENSE.