skill-creator

To create and evaluate a SKILL.md file end-to-end, you draft the initial content, organize supporting scripts, run structured evaluations, and iterate the skill description based on benchmark results to optimize triggering accuracy.

The best way to benchmark skill performance is to run automated evaluations, parse results with graders, and generate viewer-ready benchmarks to compare variants, which guides data-driven improvements to prompts and descriptions for better trigger accuracy.

The skill description optimizer works by tracking history and suggesting wording changes during iteration, directly measuring and refining variants to maximize triggering accuracy based on automated evaluation results.

Yes, you can use pyyaml to organize optional resources like scripts, references, and assets, which support deterministic tasks, in-depth information, and output templates within your skill creation workflow.

You should run structured evals during skill creation to identify edge cases, benchmark changes across iterations, and validate that your workflow and prompts perform correctly before finalizing the skill description.

Limitations to automating skill description optimization include the need for multiple iterations to stabilize triggering performance, requiring you to continually perform evals, review results, and refine wording until accuracy stabilizes.