Build a dark factory that’s right for you

A dark factory for code: an autonomous, multi-agent build crew in Claude Code.

I describe what I want in plain language, and a crew of Claude agents does the labour — designing, building, testing, and double-checking — with a single human yes/no before anything lands on main. Work is batched into waves: each wave splits into tasks, and every task runs a fixed five-stage cycle — brainstorm → implement → test → review → wrap-up — as its own Claude Code subagent in its own git worktree, so many build in parallel without colliding. The factory runs each agent as a full interactive Claude session in a tmux pane (rather than headless claude -p), signalled complete by a Stop-hook; I drive and watch the windows through cmux. The whole run is journaled to an on-disk ledger, so it survives reboots and resumes where it left off.

The quality net is a Claude Code skill I lean on constantly — a talmudic study, where N orthogonal expert agents each write an independent position on a design, then argue it out (filing challenges, conceding points, revising) until a redactor synthesizes one settled design with the dissent preserved. The factory runs two of these panels over every finished wave and auto-spawns corrective tasks for anything they catch. Then it parks the wave at a “front desk” and asks the only question I ever have to answer: merge, or hold? I’ll show how it’s wired and share the two pieces that do the most work — the talmudic-study skill, and the plain-language map of the factory itself.

Two things are shared here. Both are deliberately product-agnostic — the patterns work on any codebase.