skill-creator

To create a Claude Code skill that triggers reliably, draft the SKILL.md content, design eval prompts, and iteratively run trigger evaluations using benchmark feedback to optimize the skill description and reduce false triggers.

Trigger evaluation and benchmarking is the process of generating test prompts, executing comparisons, and measuring skill performance to ensure Claude invokes the skill when needed and avoids false or missed triggers.

Your Claude Code skill over-triggers or under-triggers because the SKILL.md description lacks precision. Running qualitative evaluations and quantitative benchmarks helps identify variance and rewrite the trigger description.

To improve existing skill triggering accuracy without starting over, apply description optimization by tuning the SKILL.md text, running held-out query evaluations, and looping improvements based on quantitative benchmark results.

The best way to troubleshoot Claude Code skill discovery behavior is to run an iterative improvement loop using qualitative feedback and quantitative benchmark results to rewrite and optimize the skill description.

You do not need to write eval prompts manually to test skill performance. The skill creation workflow designs eval prompts automatically, executes comparisons, and measures performance using generated test sets.