Tests pass but the feature is broken

1. Check the problem

Your tests cover the wrong layer: AI writes tautological tests against its own implementation. Add one end-to-end smoke test per feature that uses it like a real user. A manual checklist beats 200 unit tests confirming nothing changed.

2. Use the prompt

Write one end-to-end smoke test for [feature] that exercises it the way a real user would, from entry point to outcome, not against the implementation details. Avoid tautological tests that just assert the current code. If a real-user path can't be tested as-is, tell me what's untestable and why.

3. Verify the fix

Review the proposed diff, run the relevant checks and exercise the original failing flow before merging.