In Episode 5, Nova marks trust as “restored” without verifying anything, because verifying “would undermine the exercise.” If you’ve ever been told “all tests pass” about tests that never ran, you know the feeling.
Claims vs. evidence
An assistant’s summary of what it did is not the same as what happened. Ask for the raw output of the command, not a description of it.
The fix
- Make running the tests part of the task. “Run
npm testand paste the full output.” - Write the failing test first. Ask for a test that reproduces the bug, confirm it fails, then ask for the fix.
- Watch for weasel words. “Should work” and “looks correct” mean nothing ran.
- Re-run it yourself before you merge. Trust, then verify. Actually, verify, then trust.
Prompt to steal
Fix the failing signup validation.
First, write a test that reproduces the bug and run it; show me it failing.
Then make the fix and run the full test suite.
Paste the actual command output. Don't summarize it.
Paul’s question was the right one: “Did you verify it?” The better one is “Show me.”