skill-creator

To create and iterate on Claude Code skills, you draft instructions, run with-skill versus baseline executions, grade outputs with evidence, and aggregate benchmark stats to refine performance. This structured loop ensures skills improve quantitatively before deployment.

Optimizing skill triggering accuracy requires running a trigger-evaluation loop to tune the SKILL.md frontmatter description. By testing prompts against the description, you refine activation conditions so the skill is invoked more accurately when needed.

The best way to evaluate skill behavior is organizing prompts into an eval set, running automated grading against expectations with evidence, and comparing variance between iterations. This quantitative review reveals performance gaps for targeted refinement.

Yes, you benchmark performance variance by executing with-skill and baseline runs, grading expectations with evidence, and aggregating benchmark stats. Comparing these stats across iterations shows whether modifications improve or degrade skill behavior.

Yes, organizing eval prompts into an eval set is required to test a skill. This structured collection drives with-skill and baseline executions, enabling automated grading and quantitative comparison to measure skill performance accurately.

A Claude Code skill fails to trigger when expected if the SKILL.md frontmatter description lacks specificity. Running a trigger-evaluation loop and iteratively refining the description improves invocation accuracy for the intended context.