skill-creator

To create and test Claude skills, use a structured workflow to draft a SKILL.md file with frontmatter, write eval prompts, run parallel multi-round evaluations, and synthesize improvements from the collected results.

Automated evaluation works by running parallel trainer and test prompts against your drafted skill, collecting quantitative metrics to measure performance, and generating benchmark reports to guide iterative improvements.

The best way to structure SKILL.md is to write a concise YAML frontmatter for metadata and triggering descriptions, followed by detailed instructions, expanding the content with optional references and scripts as needed.

Yes, you need the 'anthropic' and 'pyyaml' Python dependencies installed to manage YAML frontmatter parsing and integrate with the evaluation framework for running Claude skill benchmarks.

Yes, you can optimize skill descriptions by using the built-in evaluation loop, which analyzes the quantitative metrics from test runs to automatically refine and improve the triggering descriptions in your SKILL.md.

Manually refining Claude skill prompts lacks systematic capture of intents and multi-round evaluations, leading to slower iteration cycles compared to an automated loop that synthesizes improvements from benchmark reports.