Check reference / Verifiability & Maintainability
VERIFY-01 Evals or tests present
Deterministic — runs on every scan, no LLM or network needed.
Why it matters
Without evals there is no way to prove the skill does what its description claims — every edit is a blind change, and regressions ship silently. "It worked when I wrote it" is not evidence.
How to fix it
Add an evals/ (or tests/) folder with 2–3 test prompts and their expected outcomes. The ladder of proof: runnable eval/test files pass; a prose "## Verify" section in the body only warns (manual verification); neither fails.
Example
✓ passes
evals/convert-invoice.md:
Prompt: "Convert invoice.pdf to markdown"
Expect: a Markdown table with one row per line item; totals match the PDF.