What problem does it solve?
Reduce the friction of designing, testing, and iterating on Claude-style skills by providing a repeatable, well-documented workflow that covers drafting SKILL.md, creating evals, running with-skill vs baseline comparisons, grading outputs, and packaging the final skill. It saves time and improves reliability by bundling scripts and conventions so authors can focus on intent and evaluation rather than orchestration.
Core Features & Use Cases
- Guided Skill Authoring: Templates and guidance for SKILL.md frontmatter, trigger descriptions, and instructional bodies so new skills are discovery-ready.
- Eval + Improve Loop: Tools to create eval sets, spawn parallel with-skill and baseline runs, capture timing/tokens, draft assertions, and iterate descriptions based on held-out tests.
- Benchmarking & Review: Aggregation, viewer generation, blind comparison and analysis agents to produce benchmark.json, human-review HTML, and actionable recommendations.
- Packaging & Utilities: Validation and packaging scripts to produce distributable .skill files and helper scripts for common patterns (grading, aggregation, report generation).
- Use Case: Create a new skill, run 2–3 realistic test prompts, compare with baselines, iterate the description, and produce a packaged .skill ready for distribution.
Quick Start
Ask the assistant to draft a SKILL.md for a new skill called my-skill, produce 2–3 realistic eval prompts, run the evaluation-and-improve loop, and return a packaged .skill for installation.
