skill-creator

To create Claude skills, draft your concept and compose a SKILL.md file with valid front matter, a name, and a description. You can then use evaluation tools to test triggering and iterate on the skill body.

You can evaluate Claude skills by running run_eval and interpreting results with the eval-viewer and benchmark tools. This allows you to compare configurations, analyze feedback, and iterate until skill triggering is fully optimized.

The best way to iterate on a skill is to run initial evals, review the results using the benchmark and eval-viewer tools, refine the SKILL.md body or bundles, and re-run evaluations to verify improved triggering performance.

Yes, YAML is required to structure the front matter in SKILL.md files. Valid front matter is necessary for the skill-creator suite to correctly capture intent, define triggers, and process the skill configuration.

Yes, you can include optional bundles in scripts, references, and assets directories. These bundles support deterministic task automation and store reference materials needed to enhance the skill's core functionality.

The main limitation is that SKILL.md requires valid front matter with at least a name and description to function. Without these required fields, the skill-creator suite cannot parse the file or run evaluations.