SkillsBench is a benchmark, introduced in SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks, for measuring how well Agent Skills — structured packages of procedural knowledge that augment LLM agents at inference time — actually help on agentic, expertise-heavy work.
Contents
The benchmark’s current inventory, as described in the paper, contains 87 tasks across 8 domains. Each task is paired with curated Skills and a deterministic verifier, which lets the same task be run with and without Skills and scored the same way in both conditions.
Use
SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks runs the 87 tasks under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. It reports that curated Skills raise the average pass rate from 33.9% to 50.5%, with configuration-level gains from +4.1 to +25.7 percentage points; that focused Skills with at most three modules outperform larger or exhaustive bundles; and that smaller models with Skills can match larger models without them.