We built a factory, not an app. Five Claude Code seats in Band Desktop share one room and one Git repository, and the task message is the only human input. The Coordinator plans and routes work but never writes code. The Spec Auditor turns every sentence of each stage specification into a numbered requirement, lists the edge-case hazards (races, retries, time zones, upgrades) and writes black-box acceptance tests from the spec. That way the band builds to the specification, not to the partial checks it can see. The Backend and Frontend Engineers each take one scoped work item at a time and hand off the full commit id. The Reviewer checks out that exact commit in a clean worktree and builds it with no runtime network. It runs every suite so far plus the event checker in isolated mode, looks at every screen at phone and desktop width, and returns ACCEPT or REJECT with evidence. A REJECT goes back to the owner through the room until the stage closes. In the submitted run the factory reached stage ⟨N⟩ of 4 in ⟨X⟩ hours for ⟨$Y⟩ of model spend, with ⟨Z⟩ reviewer rejections that changed the work. Before the judged run we rehearsed the whole loop on the practice track, where all four stages were accepted and every suite passed. We fed each failure from that rehearsal back into the mandates. Every seat's mandate is generic: point it at a different specification and it still applies. FACTORY.md explains how to stand the factory up, what each design choice cost, and how it catches bad work.
Category tags: