skill-creator

To create a Claude skill from scratch, you draft a skill idea, write its SKILL.md file, run structured evaluation loops, and iterate until triggering is reliable. This workflow guides capturing user intent and packaging the final skill.

Evaluating triggering accuracy involves running structured experiments and benchmarks against your existing skill. You iterate with automated improvements and evaluation prompts to measure coverage, performance, and triggering reliability.

The SKILL.md file defines the skill's metadata and processing logic. It captures user intent and acts as the core configuration file to implement, test, and package Claude skills within a repeatable workflow.

Yes, YAML is required as a dependency to draft and package Claude skills. It structures the skill metadata and configuration needed to implement the workflow and run evaluation scripts.

The best way to iterate on a Claude skill is running structured evaluations to measure triggering accuracy, then applying automated improvements and benchmarks. You repeat this loop until performance is reliable before packaging.

Yes, you can modify an existing Claude skill by updating its SKILL.md and running the evaluation workflow. This applies structured experiments and benchmarks to optimize triggering accuracy without starting from scratch.