skill-creator

To create and refine Claude skills, draft a SKILL.md file with the correct frontmatter and instructional body, then run an evaluation loop to iteratively improve triggering accuracy and output quality based on metrics.

The workflow for evaluating prompt accuracy involves running prompts through an evaluation loop, gathering results, and automatically proposing improvements to refine the skill based on feedback and performance metrics.

Description optimization for skill triggering works by using an optional module to iteratively improve the triggering description, ensuring the skill activates correctly when loaded by Claude.

Yes, you need pyyaml to parse and validate SKILL.md frontmatter, ensuring consistency and correct formatting before running the evaluation and improvement scripts.

You can automate the skill creation process by leveraging repository scripts like run_eval, improve_description, and aggregate_benchmark to orchestrate evaluation loops and analyze benchmark results.

The limitations of using automated eval loops for skill creation include relying on the quality of initial SKILL.md drafts and depending on specific repository scripts to validate frontmatter and aggregate benchmark data.