Skill Creator

This one handles the full lifecycle: scaffolding new skills from scratch, running eval loops with test prompts, benchmarking performance with quantitative metrics, and optimizing skill descriptions so they actually trigger when they should. The workflow is interview the user about intent and edge cases, draft the SKILL.md, generate test cases, run Claude against them, review results qualitatively and quantitatively, then iterate. It also includes a separate flow for tuning descriptions since undertriggering is apparently a common problem. Most useful when you're serious about testing rather than vibing, though it explicitly supports both modes depending on your tolerance for ceremony.