Skill Testing

Prove a skill still does what its SKILL.md says — by running it blind and checking the result mechanically. The discipline is the whole point: execution and verification never share knowledge.

Step 1 — Identify the target

Collect: the target skill name/path, the input (URL, file path, transcript, topic), and the output location if it writes files. Ask only if these are unclear.

Step 2 — Execute with a minimal-prompt subagent

Dispatch a fresh-context subagent (in Claude Code, the Agent tool). Its prompt contains ONLY: the target skill reference, the input, and the output path. See references/execution.md for the exact template and sandbox setup.