Claude Code verification rules: differential tests, REVIEW.md isolation, and mined rules
minchoi · x · 2026-11-19
minchoi's thread continues with rules 4-7 from the Claude Code team: run old and new systems on the same inputs for differential testing during migrations; hand off to humans after 3 failed fix rounds; never let a change edit the check that judges it (REVIEW.md read from base branch, cheap checks first); mine human review comments into rules after 2+ PRs, with 3 agents trying to disprove; flaky tests are banned from skipping, sleeps, and retry-until-pass.
More from coding & agent
- Claude Code rules: 3 failed fixes then a human, and checks a change can't edit — minchoi · 2026-11-19
- Property tests for invariants: throw thousands of generated inputs at rules that must never break — minchoi · 2026-11-19
- Give Claude a way to run the change, not just the tests: verify skill in .claude/skills/ — minchoi · 2026-11-19
- Claude Code team says Claude writes and reviews ~90% of their code, aided by 7 verification rules — minchoi · 2026-11-19
- Replit MCP Server lets AI assistants manage workspaces, files and deployments — modelcontextprotocol · 2026-11-19
- Developer flags Codex background agent reading internal chats and acting without authorization — HarveenChadha · 2026-11-19