skill-creator

Creating Claude-style skills involves defining intent, drafting a SKILL.md file, and running iterative evaluation loops to measure performance and refine prompts. This workflow provides a structured approach to capture intent, establish evaluation criteria, and optimize triggers.

A SKILL.md file requires frontmatter with a name and description, along with optional bundles for scripts, references, and assets. Structuring it correctly ensures clear trigger criteria for user prompts and integrates toxicity and safety checks during metadata generation.

Iterative evaluation loops improve skill triggering by running benchmarks, analyzing variance, and rewriting skills based on feedback. Using the description optimization workflow refines trigger reliability, ensuring the assistant activates accurately for specific user prompts.

Yes, editing existing skills and benchmarking performance across multiple prompts is supported. The workflow allows you to analyze evaluation results, apply feedback, and run variance analysis to iteratively improve the skill's behavior and response accuracy.

Toxicity and safety checks are required as part of the metadata generation process. These checks ensure that the newly created or edited assistant skills comply with safety standards before entering the iterative evaluation and refinement phase.