Agent Skills are mostly misused: don't ask a model to write its own skill, fill the gaps it can't see

You're probably using Agent Skills wrong

Anson Biggs critiques common Agent Skills mistakes, citing the SkillsBench paper. The benchmark covers 86 tasks across 11 domains with 7 agent-model configs. Curated Skills lift average pass rate by 16.2 pp, but the spread is wide: +4.5 pp for software engineering, +51.9 pp for healthcare, and 16 tasks show negative deltas. The paper's self-generated Skills condition—prompting the model to write procedural knowledge before solving—shows no benefit on average. Biggs calls this a reinvention of thinking blocks that misses the model's real knowledge gaps. His fix: after the agent gets stuck, ask what gap kept it from solving the task, then write a Skill to fill that gap. Also use Skills for repetitive project-specific workflows to save tokens. He says he edited the benchmark to use his approach and got strong results, but the post does not disclose the exact pass rates.

Why it matters: A practice-oriented critique backed by benchmark data, not empty opinion. Hits all three HKR axes, but it's a personal blog synthesis rather than original research or a product launch — scores at the featured threshold of 72. Only the excerpt is available; full argument streng...