skill-creator

To design and evaluate agent skills, draft your SKILL.md file, run structured trigger-evaluations with train and test splits, review the results, and iterate on the description until performance stabilizes for deployment.

The best way to optimize skill descriptions for consistent trigger behavior is to run trigger-evaluation workflows, analyze the generated evaluation reports, and iterate on the SKILL.md metadata based on those structured results.

You package validated agent skills by running local packaging and validation runners after your evaluation iterations reach a stable description, ensuring safe deployment and consistent trigger behavior for the final skill.

Yes, you need PyYAML installed, as it is the required dependency for running the local tooling that handles skill metadata parsing, validation, evaluation workflows, and final packaging operations.

Yes, you can benchmark an existing skill by running the trigger-evaluation workflows with train and test splits directly on your current SKILL.md, generating reports that guide targeted description optimization without a full rewrite.

Limitations when running train and test splits for skill evaluation include relying entirely on local tooling execution for validation and packaging, which requires proper local environment setup to accurately generate structured results and improvement suggestions.