I Just Deleted All My Agent Skills

The case for deleting most AI-agent skills rests on evidence that generic instructions often add little or actively hinder coding agents. In SWE-Skills-Bench, 92 of 98 publicly available software-engineering skills produced no significant change; the average task pass ratio moved only from 76.6% to 76.9%. A web-development benchmark found public skills slightly reduced success while increasing token use by 72% to 394%. AGENTS.md instruction files likewise showed no general improvement in task completion and raised costs by more than 20%. The concern is that extra directions can constrain a capable model, consume context, and make work more expensive without improving results.

That does not mean all skills are useless. SkillsBench found that carefully selected, curated skills raised average pass rates from 33.9% to 50.5%, suggesting that useful guidance depends on its quality and relevance. Instructions that merely restate documentation may be redundant as models improve, while a person’s distinctive workflow or voice cannot necessarily be inferred and may warrant explicit guidance. The case for trimming prompts is reinforced by Anthropic’s report that shortening Claude Code’s system prompt by more than 80% caused no measurable evaluation loss, and by security research indicating vulnerabilities in roughly a quarter of 31,132 skills, with another study finding similar problems. The practical conclusion is to remove generic, untested instructions, retain genuinely distinctive guidance, and evaluate whether each skill helps enough to justify its cost and risk.