skill-creator

To create a new Claude skill, you translate the task into triggers, test prompts, evaluation criteria, and improvement loops, then capture workflow intent directly into a SKILL.md file using the skill-creator automation.

Evaluation loops for prompt engineering require drafting test prompts, running with-skill and baseline tests, grading outputs, and capturing timing using helper scripts like run_loop, run_eval, aggregate_benchmark, and generate_review.

You benchmark and grade AI skill outputs by running the aggregate_benchmark and run_eval scripts to execute baseline tests, capture grading results, and generate benchmark reports for iterative description tuning.

Yes, you need Python with the pyyaml dependency installed to run the provided helper scripts that execute benchmark tests, aggregate grading results, and iteratively refine skill definitions.

Yes, you can refine an existing skill definition by applying the workflow to update triggers, run new evaluation loops against baselines, and optimize the description using captured eval metadata and viewer feedback.

You should avoid an eval-based workflow design approach if your task cannot be translated into clear test prompts, success criteria, or baseline comparisons required for meaningful grading and benchmark aggregation.