Give Claude a feedback loop

Why it matters

When an agent writes the code, a late signal turns the reviewer into the bottleneck. Generated changes are acceptable only when executable checks, not the agent’s own report, say they work.

What to practise

Wrap verification in single commands that exit non-zero on failure, and list them in CLAUDE.md with what healthy output looks like. State quantifiable targets so Claude can check the work without asking. For a bug, have Claude write and commit a failing test before the fix; for UI, iterate against a mock through a browser or screenshot tool. Then protect the loop: a hook that blocks edits to test files during a fix, or a review that rejects any diff touching a test.

How it fits

This is 7.2.3’s owned completion criteria: pasted toolchain output as evidence, enforcement through hooks, and a code owner who reviews intent and risk. Compare the in-session loop with the independent verifier the lesson distinguishes from it, and with 7.2.1’s measure of verification cost: first-pass CI success and review time per pull request.