Agent-skills-eval: Proving Agent Skills Actually Improve Model Performance

While Agent Skills from Anthropic make it easy to ship domain knowledge to models, proving they actually work is the hard part. agent-skills-eval solves this by running the same prompts twice—once with the skill and once without—then having a judge model grade both outputs side-by-side. This empirical approach transforms Agent Skills from assumption-based deployments into evidence-backed decisions, giving developers concrete metrics to validate whether their integrations truly enhance performance.