skill-creator

You can create and validate Claude skills by drafting a SKILL.md file with required frontmatter, configuring evals with graders, and running an automated evaluation loop to measure triggering accuracy and performance. The eval-viewer tool lets you compare outputs and iterate on the skill description.

The best way to improve skill triggering accuracy is through iterative evaluation. Run evals to compare outputs, use the built-in eval-viewer to quantify performance, then rewrite descriptions and expand tests based on the feedback to strengthen the skill's triggering behavior.

Yes, you need the anthropic and pyyaml Python dependencies to run the skill creation and evaluation loops. These packages support the automated eval orchestration and YAML frontmatter processing required for generating and validating SKILL.md files.

Yes, you can bundle optional resources under the scripts/, references/, and assets/ directories for on-demand use. This allows Claude to access supporting scripts and reference materials alongside the core SKILL.md documentation when the skill triggers.

The evaluation loop works by running evals with graders to compare current outputs against expected results, then viewing results with the built-in eval-viewer to quantify skill performance. You use this feedback to rewrite descriptions, expand tests, and continuously improve triggering accuracy through iteration.

SKILL.md frontmatter should capture the skill's intent, defining what it enables Claude to do and the contexts in which it should trigger. The skill creation process generates this required frontmatter along with a detailed body, reducing manual guesswork in the skill design process.