The Claude Code team, where ~90% of code is written and reviewed by Claude, relies on automated verification loops to scale parallel agent execution. Internal testing showed a separate grader checking output against a rubric improved structured file generation success by up to 10 points over standard prompting. The team prioritizes three test types: end-to-end (E2E) tests for real user flows using tools like CDP or simulators, which provide definitive done criteria; property-based tests via libraries like fast-check or Hypothesis to generate thousands of edge-case inputs that expose agent blind spots; and differential tests comparing old and new system outputs on random or production-sampled inputs for migrations. A "verify" skill in .claude/skills/ codifies E2E smoke tests, which Claude runs before committing. The review pipeline enforces local test/lint/build passes via hooks, stamps commits with verification status, and uses a commit gate with observe/enforce modes. A 39-item reliability checklist triggers targeted checks from diffs. Code Review reads REVIEW.md from the base branch to prevent rule tampering, tags findings as Important/Nit/Pre-existing, and tunes recall against an evaluation suite recreating past incidents with one-sentence answer keys. Rules derive from patterns in reverts (requiring two PRs and three-agent disproof) or replaying bug finders on incident-linked PRs. Flaky tests are deleted if root causes are unfound. Verification artifacts are routinely pruned to maintain signal.
Read original
hackernews