We built a generic software factory, not an app. Four Claude Code seats run in one BAND Desktop room. The Architect plans and sends every handoff. The Implementer writes one scoped work item at a time. The Reviewer attacks each commit against every spec sentence. The Verifier runs a no-cache build and the organisers' checks in isolated mode (no network, 2 vCPU, 2 GiB) before releasing anything. The mandates say how each seat works, never what is built, and they pass the organisers' vocabulary scan for both tracks. The product lives only in the single task we posted to @architect at 20:32:50Z. Nothing else was posted. Result: all 4 pocketful stages were released in 6 h 14 m. On a fresh clone in isolated mode the shipped checks pass 147, 35, 6 and 5 (claimed stage 4). There were 28 seat commits, each under its own seat's name, and 87 seat messages. The Reviewer issued 3 rejections that changed the code, for example a hold that could be captured for 0 and an import that pushed the clock ahead. Estimated cost: $60.09 at list prices. We document what failed: two earlier attempts stalled (an MCP name clash, then a dropped reply). Both were fixed generically before the clean run. FACTORY.md is enough to stand the factory up, and room.json is the full room.
Category tags: