Most SKILL.md files are reminders, not teaching. Nikolay Petrov of Intel told AI Engineer that frontier models already knew about 70% of what popular skills contain.
I think teams should delete before they add. Petrov showed the check is simple and brutal, and he said a no-skill run solved three of five tasks while the skill-loaded run solved all five. Do not ship the extra tokens without that proof.
Petrov broke skills into 25 to 30 testable items with paired questions and answers, then ran each with and without the skill under a judge panel, he said. That yields four buckets: already known, harmed by the skill, newly taught, and missed both ways. Harm is rare in popular work, with Petrov citing one bad item in about 400 across Anthropic skills, but he said smaller repo skills fail more.
The audit is cheap, with Petrov putting 29 skills at 14 dollars in Sonnet judge and Haiku calls. Strict recall collapses fast, and he said Opus dropped from 70% on general ideas to 22% on exact keys, values and steps. Smaller models knew less to start, while Sonnet knew more, which is why exact details still pay.
The format makes bloat easy. A SKILL.md stores name and description in frontmatter and instructions in the body, and agents read only name and description at startup then load the body on demand, per the AscendingDC explainer. Jane Street's socratic-tutor skill in its public agent_skills repo shows the opposite intent, a drill built around placement, levels zero to five, spaced review and final assessment. Petrov conceded that steering skills do not need novelty to be useful, and he is right. Novelty still wins for knowledge skills, because it tells you what to cut.
If you maintain a skill, cut what the model knows, keep the exact parameters, and rerun evals before you merge. I would not let a 70% known file burn context on every run.
