skill-creator

To create a Claude Code skill, you draft a SKILL.md file defining the workflow and best-practice structure, then iteratively refine it using an evaluation loop to ensure accurate triggering.

Your Claude skill triggers incorrectly when its description is not optimized. You can fix this by running trigger-rate benchmarking evals to measure should-trigger versus should-not-trigger outcomes and rewriting the description.

An eval-driven loop for skill description optimization is a process of generating test prompts, running trigger checks, analyzing results quantitatively, and rewriting the skill description until trigger performance improves.

Yes, you can improve trigger accuracy for an existing SKILL.md by creating test prompts, running trigger checks to review results, and iteratively updating the description to match correct user intents.

Trigger-rate benchmarking works by evaluating should-trigger and should-not-trigger outcomes against generated test prompts, allowing you to analyze results and expand test coverage over iterations.

The best way to evaluate Claude skill invocation behavior is using an eval-driven loop that quantitatively and qualitatively reviews trigger checks, expanding test coverage to refine the SKILL.md description.