Check reference / Verifiability & Maintainability
VERIFY-04 Verification step present
LLM-assisted — runs with an Anthropic key (BYOK on the CLI, managed on Pro hosted scans).
Why it matters
A skill that produces output and stops assumes the happy path succeeded — so the model ships its first attempt unchecked, and the failures land on the user instead of being caught in-session.
How to fix it
End the instructions with a concrete self-check before finishing: run the tests, re-read the output against the request, confirm the file parses — and say what to do when the check fails. An LLM judges the step (pass at 80/100); vague "double-check your work" scores low, actionable verification scores high.
Example
✗ flagged
Generate the config file and tell the user it's ready.✓ passes
After writing the config, run `app validate config.yml`. If validation fails, fix the config and re-validate before reporting done.