skill-creator

Creating and evaluating AI skills involves drafting a SKILL.md file, generating test prompts, and running paired baseline evaluations to benchmark performance. This iterative workflow uses scripts to spawn runs, capture timing, grade assertions, and aggregate results for reliable skill improvement.

Iteratively improving AI skills requires running blind comparisons between with-skill and baseline outputs, grading assertions, and generalizing fixes. By optimizing the SKILL.md description and reviewing aggregated benchmark artifacts, creators can reliably improve trigger rates and reduce output variance.

Yes, filesystem workspace access is required to author and test reusable AI skills. The evaluation workflow depends on this access to write SKILL.md files, bundle scripts, store test cases, and generate review viewers for human feedback during the iteration loop.

Comparative evaluation works by spawning paired with-skill and baseline runs to test AI skill logic. The system captures timing and token usage from both runs, grades assertion pass rates, and aggregates benchmark data to surface actionable feedback for the next iteration.

Yes, you can convert a manual workflow into a repeatable AI skill by capturing user intent to produce a draft SKILL.md. The workflow then generates realistic test prompts, runs parallel automated evaluations, and iterates until pass rates and variance are acceptable.