Skill Benchmarks

Each skill is run against a set of eval prompts twice — once with the skill injected, once as a bare baseline. An LLM-as-judge scores each assertion. The delta is the signal.