Skill Sommelier: Claude Code Plugin
A curated plugin that discovers, manages, and distributes Claude Code skills from GitHub.
Overview
A personalized self-improving Claude Code skills manager. Discovers skills from GitHub, ranks them to your profile, and uses its own skills to validate, optimize, and evolve the collection.
Problem
Finding quality Claude Code skills across GitHub is manual and time-consuming, and it is difficult to tell which skills are good and which are not.
Approach
Built a Claude Code plugin repository. It includes meta-skills for personalized skill discovery, automated skill creation and consolidation, and ongoing tuning and optimization.
Constraints
- No application code, purely text-driven
Key Decisions
Pure text architecture
For fun, and explore alternatives way of `coding` in a low risk environment.
Self-evolving meta-skills
Including skills that discover, craft, and update other skills creates a self-improving ecosystem without manual curation.
Claude Code plugin marketplace distribution
The marketplace provides a standard install path, making adoption frictionless for Claude Code users.
Benchmark-driven skill tuning inspired by karpathy/autoresearch
Manual prompt tuning is guesswork. A fixed-evaluation iterative loop turns skill quality into a measurable optimization problem.
Result & Impact
Streamlined my own Claude Code workflow by providing reusable skills for common tasks like Python project setup, slide creation, and repository simplification.
Learnings
- A text-only architecture can be surprisingly powerful when the consumer is an LLM.
- Meta-skills that improve the skill collection itself create a compelling self-evolving loop.
- Borrowing the fixed-eval iterative loop from ML training (à la autoresearch) works surprisingly well for prompt optimization.