skill-creator

To create and test Claude skills, you draft a SKILL.md file with required name and description frontmatter, generate test prompts, and use Python tooling like run_loop.py to execute with-skill vs baseline evaluations and grade results for benchmark reports.

A SKILL.md file defines Claude skills and requires valid frontmatter with a name and description. It optionally bundles scripts, references, and assets to provide structured instructions and resources for the skill.

You benchmark Claude skills by using run_eval.py and aggregate_benchmark.py to spawn with-skill and without-skill baseline runs, grade outputs, and aggregate metrics into human-reviewable reports for comparison.

Yes, running Claude skill evaluations requires Python with the anthropic and pyyaml dependencies installed to execute the included tooling for generating reviews and aggregating benchmark metrics.

Optimizing Claude skill descriptions for trigger accuracy involves running iteration loops with run_loop.py, analyzing blind comparisons, and applying reviewer workflows to refine the SKILL.md description field.

Yes, after running evaluations and generating benchmark reports, you can package validated Claude skills by bundling the SKILL.md with optional scripts, references, and assets for distribution.