skill-creator

To create a GLM code skill, you scaffold a SKILL.md file with frontmatter, run trigger evaluations using train and test prompts, and iterate on the description until triggering accuracy improves across configurations.

Evaluation-driven iteration runs prompts against your skill, collects performance metrics, and applies description optimization to refine triggering accuracy, generating benchmarks to track progress across multiple iterations.

Yes, pyyaml is a required dependency for running the skill creation workflow, as it parses the SKILL.md frontmatter and supports configuration loading during the evaluation and benchmarking process.

The best way to track skill performance is generating benchmarks after running trigger evaluations, which provides reporting metrics to compare results and monitor progress across multiple iterations and configurations.

Yes, you can use the built-in description improver to optimize an existing SKILL.md, leveraging evaluation metrics from test prompts to refine the text and increase triggering accuracy and clarity.

Inconsistent triggering usually indicates the SKILL.md description needs refinement; running trigger evaluations with a mix of train and test prompts identifies gaps to iterate and improve accuracy.