Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-08 · high
The method found most of what's in a typical skill file never changes what the agent does, while some reference files add tens of thousands of tokens a call.
Why now: This lands as more builders write SKILL.md files for agents that get called over and over, exactly the repeated-cost pattern the paper is measuring.
SkillReducer strips more than half the prose out of a typical agent skill file and still keeps full capability, according to new research posted to arXiv.
More than 60% of what's written in a typical SKILL.md file never changes what the agent actually does when it runs, the paper found. Some reference files tacked onto skills quietly inject tens of thousands of extra tokens into every single call.
The fix isn't a careful rewrite. Mechanical compression, cutting narrative and pushing detail into on-demand files, trimmed tokens by a mean of 77.5% across the tested skills. Capability didn't drop.
That's a cost most people writing agent skills haven't priced in. Every SKILL.md file gets read on every relevant call, so extra prose is a recurring token cost, not a one-time file-size problem. If I'm writing instructions an agent will use repeatedly, the backstory and caveats belong in a file it loads on demand. The main file should hold only what changes behavior.
The paper doesn't say how many distinct skills it tested this compression against, or whether the 77.5% figure holds for skills with heavy branching logic. But the direction is clear: write the skill file for the model that executes it, not for the person reading it once.
Each link below shares sources, entities, or timing with this story.
Shared entity: Skill / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Skill; reported by the same outlet (arxiv.org); overlapping topics (agent, token).
Shared entity: SKILL / Same source domain / Shared topic / What happened next
Both cover SKILL; reported by the same outlet (arxiv.org); overlapping topics (agent, doesn).
Both cover SKILL; reported by the same outlet (arxiv.org); overlapping topics (agent, call).
Shared entity: SKILL / Same source domain / Shared topic / Earlier coverage
Both cover SKILL; reported by the same outlet (arxiv.org); overlapping topics (agent, behavior).
Shared entity: Skill / Same source domain / Shared topic / Earlier coverage
Both cover Skill; reported by the same outlet (arxiv.org); overlapping topics (call, capability).
Shared entity: Skill / Same source domain / What happened next / Downstream implication
Both cover Skill; reported by the same outlet (arxiv.org); picks up the Skill thread on 2026-08-09.
Shared entity: Skill / Same source domain / What happened next / Tension
Both cover Skill; reported by the same outlet (arxiv.org); picks up the Skill thread on 2026-08-06.
Both cover Skill; reported by the same outlet (arxiv.org); picks up the Skill thread on 2026-07-28.