Build Harness v2

This is the harness I build with now. It's two skills, a reviewer subagent and a few rules, sitting on top of stock Claude Code. There are no commands to remember - you describe what you want and the right skill picks it up.

Say "I want to build …" and start-product asks a few questions, checks any outside data the idea depends on, writes a brief, a decisions log and a backlog of small slices, then builds the first slice end to end. In an existing project, say "add …", "fix …" or "next" and ship-feature checks the request against the real data source, writes failing tests, builds the change on a new branch, gets two independent reviews, tries the change in the running app, and stops. Every run ends in one of four verdicts - Ready to merge, Don't merge, Question or Blocked - with the problem in plain words. It never merges, pushes or deploys. That's your call.

It's this lean because the bigger version didn't earn its keep. I started with 16 agents and pared it down until it was exactly what was necessary - which is this. I compared the earlier harness with plain Claude Code on seven features of one real app. Correctness was the same in both. The harness cost about 1.6 times as much and took about 3.6 times as long. The one component that caught real faults was a reviewer reading the change in a fresh context, so that stayed.

The harness is open source and MIT licensed. I'd love to hear your feedback: contact@henryguyver.com