Cut Skill Bodies 39%, Task Success Rises 2.8%

Notes on SkillReducer: Optimizing LLM Agent Skills for Token Efficiency (arXiv:2603.29919) — Yudong Gao, Zongjie Li, Yuanyuanyuan, Zimo Ji, Pingchuan Ma, Shuai Wang · March 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

Agent skills are supposed to make agents cheaper and more capable, not the reverse. An audit of 55,315 publicly available skills finds the opposite happening at scale. 26.4% carry no routing description at all, and another slice sit under 20 tokens — both break or blunt the mechanism an agent uses to pick the right skill before it even reads the body. Once a skill is selected, the body doesn't fare much better: only 38.5% of its content qualifies as actionable core instruction. The rest is background prose and examples the agent didn't need, plus templates it never used. Reference files make the problem worse — a 100-skill sample alone carries 1.67M tokens across 505 files, all of it inherited on a single invocation whether or not the task touches it. SkillReducer's answer isn't a blunt compressor but a two-stage pipeline. Stage 1 rewrites routing descriptions, and generates them where missing, using delta debugging against a simulated routing oracle plus real-agent validation. Stage 2 classifies body content into core rules, background, examples, templates, and redundant material, then applies progressive disclosure: everything but the core rules moves into modules loaded only on demand, gated by a faithfulness check and a feedback loop that restores content if cutting it hurts task performance.

The numbers land where the method aims. Description length drops 48% on average (median 59%), body length drops 39% (median 43%), and total input tokens fall 26.8% end to end. What's notable is what happens to quality under that much cutting: it goes up, not down. Scored against the uncompressed originals, 25.3% of compressed skills perform better on task-based evaluation against 14.0% that regress, netting a 2.8% overall improvement — and on the independent SkillsBench benchmark, all 87 tasks hold their result with no regression at all. Call it a less-is-more effect: cutting non-essential content removes what distracts the model inside its own context window, not just what it costs to run the prompt. And it isn't a one-setup fluke. Mean retention across five models spanning four families comes in at 0.965; compression still holds around 0.944 retention when run through an independent agent framework outside the original test harness. SkillReducer beats perplexity-based pruning, direct LLM compression, plain truncation, and random removal at matched token budgets.