Progressive disclosure in Agent Skills: use exactly one level — a second routing level never helps and sometimes breaks accuracy

arXiv 2607.17598·high signal

A controlled study (arXiv 2607.17598, submitted 2026-07-20) compared raw-document navigation, Agent Skills packs with progressive disclosure at varying depths, and a classical hybrid retriever across InfiniteBench, three agent harnesses, and three model families. One level of progressive disclosure degraded gracefully and won as the corpus scaled to many books, but a second deeper routing level never helped and sometimes reduced accuracy; on single documents with a strong harness the gain was near zero. The practitioner takeaway inverts common SKILL.md dogma: deeper nesting of reference files is not free structure, and progressive disclosure 'buys context, not intelligence' — only reach for it when the corpus outgrows what the agent can read directly.