What they did
Hugonnot runs tickets through a Claude Code skill that chains five steps: plan, code, proof, review and wrap-up. One sentence picks the models: a Fable agent writes the plan, a Sonnet agent writes the code from a brief, and Opus agents review in a loop without knowing who wrote the change. The ticket was to let the author's contrast checker read two colours from the page URL and keep the URL in sync. Tests came first with Playwright in a real browser and failed for the right reason before any code existed. Hugonnot then undid the fix in several ways to check each test turned red again, and took screenshots at three screen sizes. The review loop stops when a round has no blocking or important finding and no code change, or hands control back to Hugonnot after five rounds.
How it went
Round 1 reproduced a real bug in WebKit: rewriting the URL on every colour picker event hit Safari's limit of 100 rewrites in 10 seconds, so the URL froze on an old colour. The fix was a 250 ms delay. Round 2 showed no test actually checked that delay; spacing the test's events 16 ms apart fixed it. Round 3 found nothing and the loop stopped. Nine findings in total: 5 fixed, 1 rejected, 3 deferred. The author gives no time or quality comparison against working without the setup.
Worth knowing
The rule came from an earlier change that was reviewed 15 rounds in a row because nobody set a limit and the code kept growing. Hugonnot also found that the read-only reviewers could edit files: an instruction in the prompt is only a request, so restrict tools in the agent definition.