skill-creator

To create Claude skills, you draft initial SKILL.md content from project context, then refine them through an iterative evaluation loop that runs controlled prompts, collects performance metrics, and rewrites descriptions to improve triggering accuracy.

Evaluating LLM skill triggering accuracy involves running controlled prompt evaluations, comparing baseline performance with and without the skill, and analyzing feedback to optimize the description for maximum recall across different contexts.

Yes, you can automate the iteration workflow by running evaluation scripts that capture metrics, analyze results, and automatically propose changes to rewrite the skill body and description until deployment readiness is achieved.

Yes, pyyaml is required as a dependency to parse configuration files and manage the structured data needed to generate, evaluate, and refine the SKILL.md content during the iteration workflow.

Claude skills fail to trigger when descriptions lack sufficient recall across contexts, which you resolve by running description optimization to analyze evaluation results and rewrite triggering text for accurate context matching.

You optimize skill descriptions by running controlled performance evaluations, capturing metrics on triggering behavior, and iteratively rewriting the description text to maximize recall and ensure correct triggering across diverse user contexts.