skill-creator

To create and test Claude skills, you draft a SKILL.md file with frontmatter and optional directories like scripts or assets. The Skill Creator provides a structured workflow to draft, evaluate, benchmark, and iteratively improve skills.

Evaluation and benchmarking for Claude skills involve running prompts and collecting metrics to compare configurations, such as with_skill versus without_skill. This process measures reliability and identifies areas for iterative improvement.

You optimize skill triggering descriptions by running a dedicated optimization pass that proposes updates based on evaluation results and history. This maximizes triggering accuracy and ensures reliable skill activation.

Yes, YAML is required as a dependency to manage SKILL.md frontmatter. The frontmatter loads metadata, executes deterministic tasks through scripts, and provides context during skill activation.

The best way to iterate on skill configurations is to use evaluation results and benchmark history to automatically propose description updates. This iterative improvement loop refines the skill toward a polished, trigger-optimized state.