What they did
The author is building Refmade, a desktop app that draws Claude Code and Codex sessions as an office of teams, and has Claude Code and Codex agents write most of the code while they plan and review. For the first local version they wrote a build contract in the repo, a markdown file that fixed the interfaces and gave each of five builders its own files: log classification, the Codex log reader, the office model, the local HTTP server, and the browser page. An integrator could touch anything, but only to fix defects found when joining the work. Other builders' files were read-only, so changing an interface meant editing the contract and telling the integrator. Existing tests were frozen and new dependencies were banned. After integration they ran a review from four angles and a second round that tried to refute each finding.
How it went
Findings that survived refutation became 25 fixes. The first version then ran on 127.0.0.1 with 224 passing tests, a first load of about two seconds and no outside requests in browser checks. A request for a more game-like look produced three skins on the same data and they picked one, bringing tests to 245. The workflow's own critique and fix steps failed when they hit a weekly usage limit, so they checked those parts by hand. Some timing tests were flaky under load.
Worth knowing
The author kept only the count of the 25 fixes, not what they were. Agents will ship words that only make sense inside the contract: a label they picked for workflow branches read oddly and had to be removed. No usability test with a non-developer was done, and a finished request in the app means the turn ended, not that the work succeeded.