ag-criar-skill

To create and refine Claude skills iteratively, you draft a skill with clear frontmatter metadata, run quantitative evals to measure trigger accuracy, and optimize the description through a history-aware loop to prevent overfitting.

You benchmark skill triggering by running the included harvest and grade scripts to evaluate descriptions against real prompts, comparing performance metrics across different configurations to maximize activation accuracy.

The best way to update a skill's description is using the improve workflow, which grades current performance with quantitative evals and applies targeted description optimizations to enhance prompt triggering.

Yes, you can run quantitative evals on new skills created from scratch by using the harvest and grade scripts to measure baseline triggering effectiveness and drive iterative improvements.

Yes, the skill creation workflow prevents description overfitting by using a history-aware loop that tracks previous iterations and benchmarks, ensuring optimizations generalize rather than memorizing eval prompts.

Yes, you need pyyaml and anthropic dependencies to run skill evals and benchmarks, as the workflow relies on these libraries to parse frontmatter metadata and execute the evaluation grading scripts.