I've written two QA skills for our app: 1) flow discovery, and 2) e2e implementation.
The first one scans the app and detects critical flows, happy paths, unhappy paths, invariants. Writes an invariant matrix across app domains and surfaces, and maps acceptance criteria to each.
The second actually implements the flows as e2e tests (Playwright for web, Maestro for mobile). Tests are run locally before a PR is submitted, with a video recording. The video + screenshots are posted in the PR body. A rule in CLAUDE.md says "No PR w/o visual evidence!".
CI runs the suite depending on touched code.
It's nice to lay back and watch the runner click through your app.
source & further reading
news.ycombinator.com — original article
Ask HN: How do you break AI agents out of the "apologize and repeat" loop
Show HN: Building AI agents client-side JavaScript
Ask HN: Why don't LLM APIs have a first-class test mode?