What problem does it solve?
Reduces the friction of designing, testing, and improving Claude skills by providing a repeatable workflow for drafting SKILL.md, generating test prompts, running with-skill vs baseline evaluations, grading results, and iterating until the skill meets quantitative and qualitative criteria.
Core Features & Use Cases
- Drafting & Scaffolding: Help the user capture intent, produce a SKILL.md frontmatter and operational instructions, and recommend triggering phrases and output formats.
- Eval Design & Execution: Create eval sets, spawn parallel with-skill and baseline runs (or run serially in Claude.ai), capture timing/tokens, and save iteration workspaces.
- Grading & Benchmarking: Use grader and aggregator scripts to produce grading.json and benchmark.json, surface pass rates, timing, and token deltas, and generate an HTML review for human feedback.
- Iterate & Optimize: Propose concrete rewrites, bundle reusable scripts, optimize the SKILL.md description to improve triggering accuracy, and optionally package the skill.
- Use Case: Ideal when a user wants to create a new automation skill, improve an installed skill's reliability, or rigorously compare versions with blind comparisons and benchmark analysis.
Quick Start
Tell the assistant what the skill should do, ask it to draft SKILL.md and 2–3 realistic test prompts, run the eval loop, review the generated benchmark and feedback, and then iterate until the results are satisfactory.
