skill-creator

To create Claude skills, you draft a SKILL.md file with required frontmatter, generate test prompts, and run parallel with-skill versus baseline evaluations. This process iteratively improves the skill body until Claude triggers predictably and benchmark pass rates are satisfactory.

Automated skill evaluation runs parallel with-skill and baseline comparisons, captures timing and token metrics, and grades outputs to aggregate benchmark reports. This benchmarking pipeline surfaces pass-rate, time, and token deltas to quantify skill performance improvements.

You package a Claude skill directory by using included packaging utilities to produce a distributable .skill artifact. This occurs after drafting SKILL.md, running evals, inspecting the eval-viewer, and finalizing the optimized skill description for correct triggering.

Yes, you can optimize SKILL.md descriptions using an automated loop that generates and tests improved descriptions. This description optimization increases correct skill triggering while avoiding overfitting to specific test prompts.

Yes, PyYAML is a required dependency for running the skill-creator pipeline. The Python scripts rely on it to parse SKILL.md frontmatter, process evaluation schemas, and execute grading, aggregation, and packaging flows.

The best way to evaluate Claude skill performance is running parallel with-skill and baseline evaluations using generated test prompts, then inspecting results via the eval-viewer. Grader and comparator agents aggregate benchmark metrics to confirm robustness before packaging.