Skip to content- Sources: primary
- Summary: Petrov breaks a SKILL.md into 25 to 30 question and answer items, asks the model each question with and without the skill, and has a judge from another model family score the answers. That sorts lines into already known, newly taught, harmful, and still missed. Opus understood 70% of skill content under lenient scoring but only 22% under strict scoring of exact details, and only one harmful item appeared among about 400 in Anthropic's skills.
- Why it matters: Skill authors can prune what the model already knows and keep exact details, then confirm with end-to-end evals that scores hold.