skill-creator

To build and refine Claude skills, you capture user intent, draft SKILL.md files, run test prompts, evaluate results, and benchmark performance to turn workflows into reproducible instructions. This process ensures each skill reaches production quality with measurable confidence.

Skill description optimization generates trigger queries and runs an evaluation loop to update the description so Claude invokes the skill reliably. It analyzes benchmark aggregation results to improve how the skill is triggered in agent workflows.

You set up evaluation and benchmarking by running tests with and without the skill, grading assertions, aggregating statistics, and analyzing patterns. Scripts coordinate evaluation, benchmark aggregation, and review viewing to keep the process reproducible.

Yes, pyyaml is required as a dependency. It supports the scripts used for evaluation, benchmark aggregation, review viewing, and description optimization that coordinate the iteration process for Claude skills.

The best way to capture intent is through structured interviews covering user goals, trigger contexts, expected outputs, and whether objective test cases are needed. This ensures the drafted skill aligns with the actual workflow requirements before iteration begins.

You should avoid automated skill iteration when your workflow cannot be broken into testable assertions or when qualitative feedback is unavailable for evaluation. The process requires measurable benchmark data to reliably optimize skill triggering descriptions.