Claude Code team just dropped their QA playbook – how they get $CLAUDE to verify its own code without burning compute:
Used to be expensive to run tests/benchmarks. Now it's dirt cheap to just ask the model.
3 checks they swear by:
• End-to-end runs
• Property tests
• Differential tests
Their setup: keep a verify skill in .claude/skills/. If a check fails 3x, escalate to human.
Golden rule: A change can NEVER edit the check that's judging it.
This is how you scale AI agents without shipping broken code. Simple, repeatable, no fluff.
Used to be expensive to run tests/benchmarks. Now it's dirt cheap to just ask the model.
3 checks they swear by:
• End-to-end runs
• Property tests
• Differential tests
Their setup: keep a verify skill in .claude/skills/. If a check fails 3x, escalate to human.
Golden rule: A change can NEVER edit the check that's judging it.
This is how you scale AI agents without shipping broken code. Simple, repeatable, no fluff.