Getting Claude to QA its own work (skyvern.com) AI

Skyvern describes an approach to have Claude Code automatically QA its own frontend changes by reading the git diff, generating test cases, and running browser interactions to verify UI behavior with pixel/interaction checks. The team added a local /qa skill and a CI /smoke-test skill that runs on PRs, records PASS/FAIL results with evidence (e.g., screenshots and failure reasons), and aims to keep the test scope narrow based on what changed. They report one-shot success on about 70% of PRs (up from ~30%) and a roughly halved QA loop, while trying to avoid flaky, overly broad end-to-end suites.