skill-creator

To create and test Claude skills, you can draft a SKILL.md file, generate realistic test prompts, spawn parallel baseline runs, and aggregate grading metrics to iterate on skill performance. This workflow provides a repeatable end-to-end process for authoring and benchmarking.

Optimizing skill descriptions for reliable triggering involves using an iterative optimizer to refine frontmatter and operational instructions, ensuring accurate activation. This process improves how consistently a skill is selected in interactive and Cowork environments.

Running blind comparisons and benchmarks requires spawning parallel with-skill and baseline runs, capturing timing and token metrics, and using an automated grader. A blind comparator then surfaces objective comparisons and concrete improvement suggestions.

Yes, you need Python with the anthropic and pyyaml dependencies installed to run the provided scripts for skill creation and packaging. These libraries support generating SKILL.md files, orchestrating evals, and creating distributable .skill packages.

Yes, you can use this workflow to improve existing Claude skills by running post-hoc analyzers and graders on new test cases. This surfaces quantitative metrics and human feedback to iterate until the desired skill performance is achieved.