skill-creator-claw

To create and evaluate Claude Code skills end-to-end, you can use a workflow that drafts a skill, runs evaluations against an external Agent via HTTP API, and iterates on the frontmatter and body to optimize triggering accuracy.

Benchmarking and optimizing skill triggering accuracy involves running evaluations against an external Agent, analyzing the generated eval results, and performing post-hoc reviews with an eval-viewer to track improvements across iterations.

Yes, you need an external Agent accessible via an HTTP API to measure triggering accuracy, generate eval results, and guide the iterative refinement of your skill's frontmatter, body, and workflow.

Yes, you can optimize skill frontmatter and descriptions using progressive disclosure techniques, which are supported alongside lightweight benchmarking to guide decision-making during the skill refinement process.

Without an eval-iterate loop, manual skill creation involves repetitive setup and lacks automated benchmarking, making it difficult to rapidly iterate from concept to deployed skill or track triggering accuracy improvements.