skill-creator

To create AI assistant skills, you draft a SKILL.md file with metadata and instructions, generate test prompts, run baseline and with-skill comparisons, and iterate until outputs pass assertions. This process packages workflows into reproducible, testable artifacts.

Skill benchmarking evaluates assistant prompts by generating test cases, running parallel with-skill and baseline comparisons, grading outputs against assertions, and aggregating statistics into benchmark.json. This measures performance improvements and surfaces analyst observations.

You package automation skills by bundling validated SKILL.md files, test prompts, eval workspaces, and helper scripts using provided packaging scripts. This creates reproducible artifacts satisfying grading.json and benchmark.json schemas for distribution.

Skill development relies on local utilities for YAML and frontmatter parsing, requiring only PyYAML as a dependency. This supports metadata extraction and validation within SKILL.md files during the authoring and packaging process.

Iterating trigger descriptions involves interviewing for edge cases, running eval comparisons, grading outputs, and refining metadata until the skill reliably passes assertions. This optimization ensures skills trigger correctly in target workflows.

Skill evaluation outputs fail assertions when SKILL.md instructions lack specificity, test prompts miss edge cases, or trigger descriptions cause incorrect activation. Iterative drafting, baseline comparisons, and grading against benchmark.json schemas identify and resolve these gaps.