What they did
Weese, writing on dev.to six months after his first post on the setup, gives each agent one role. A page mapper visits a page once and saves its controls, URL params, validation messages and empty states to a file in the repo. A UI planner reads that map and the ticket and never opens the browser. Other agents write manual test steps into Xray, generate page objects and specs, plan and write API tests, and repair broken locators. A generate-tests skill takes a Jira ticket or an epic key, pulls the tickets through the API, removes duplicates and writes a brief, then pauses for his approval before the plan, the manual cases and the automation. He split his large CLAUDE.md into short per-folder guides and later packaged the agents, skills and slash commands as a Claude Code plugin so each test repo can install it. Commands also file bugs and QA tasks in Jira once he confirms.
How it went
Before the page mapper, planning one feature used 156k tokens, about 78% of his context window, because the planner rescanned the page every time. Splitting CLAUDE.md cut the root file from 12,306 to 5,177 characters, roughly 1,790 tokens saved per session. In a public rebuild against a demo site, plan review caught two tests that would have passed on broken behaviour. A full run showed 19 failures; capping workers at 2 and stopping duplicate smoke runs left only the 7 that were real mismatches with the site's own docs. He says the bottleneck moved from writing tests to reading plans.
Worth knowing
Prompt rules alone did not hold, so he moved every rule he could into lint (no XPath, no fixed waits, no hidden skips). One rule stays manual: when the plan and the ticket disagree, the test must assert the ticket. The healer gets three attempts per test and then marks it test.fixme() with a note, or it keeps chasing locators that are not the real problem, such as a site returning 503 under load.