SkillReducer

Definition

SkillReducer is a two-stage optimization framework for LLM agent skills (arXiv 2603.29919, Gao et al.). Motivation: every token of skill content injected into the context window incurs both monetary cost and attention dilution — reference files can inject tens of thousands of tokens per invocation.

Key Points

• Empirical finding (55,315 skills): 26.4% of skills lack routing descriptions entirely; over 60% of body content is non-actionable; reference files can inject tens of thousands of tokens per invocation.
• Stage 1 — routing layer: compresses verbose skill descriptions and generates missing ones via adversarial delta debugging.
• Stage 2 — skill bodies: taxonomy-driven classification + progressive disclosure — separates actionable core rules from supplementary content loaded on demand; validated by faithfulness checks and a self-correcting feedback loop.
• Results (600 skills, SkillsBench): 48% description compression, 39% body compression, +2.8% functional quality — a less-is-more effect where removing non-essential content reduces distraction.
• Transfer: mean retention 0.965 across 5 models / 4 families; generalizes to an independent agent framework.

Implications

SkillReducer is the research-side proof of the “less is more” claim that token-usage-reduction and rtk-rust-token-killer dispute on the tool side: compressing skills (static instruction content) improves quality, whereas compressing tool output at runtime can degrade it. The distinction is what content is compressed: authored instructions (redundant) vs live evidence (lossy). Progressive disclosure is the same pattern as claude-code-memory’s on-demand topic files and Claude’s deferred MCP tool definitions.

Open Questions

• Whether the +2.8% quality gain holds on harder, reasoning-heavy benchmarks beyond SkillsBench.
• How the adversarial delta-debugging description generator behaves on non-English skill corpora.

Related

token-usage-reduction — the cost-lever space skills occupy per session rtk-rust-token-killer — runtime tool-output compression; the contrasting failure mode claude-code-memory — progressive disclosure analog in Claude Code claude-code-sessions — skills load per session as fixed context

Sources

• raw/external/arxiv-org-2603-29919-0aae2e40.md