# Skill RSI

Recursive Self-Improvement for Agent Skills

## What It Is

What if an Agent Skill could improve itself? Skill RSI is a recursive self-improvement system for Agent Skills, grounded in evals instead of vibes. Give it a goal and it drafts a new skill; hand it a skill you already have and it tries to make that one better. Before it writes, it researches the domain into an ontology of users, output standards, failure modes, and authorities. Then each loop treats the current champion as the control and a challenger as the treatment: form a hypothesis, mutate one focused surface, run both against the same prompts and criteria, inspect the evidence, and only promote the challenger if it wins without breaking stable behavior. It is the sequel to SkillEval: not just measuring which skill is better, but deciding what to try next and keeping the full experiment history so every pass starts smarter. Free, open source, and built to run locally as a CLI, UI, and Codex plugin.

## More Context

Skill RSI treats skill improvement like a small research lab. Before it writes, it studies the domain into an ontology: who the skill serves, what excellent output looks like, what failure modes matter, and which authorities have sharp opinions worth preserving.

Each loop has a control and a treatment. The current champion skill runs against the same prompts as the challenger, judged against the same criteria. Only one focused surface changes at a time, so a win means something.

The useful part is the evidence review. From Codex, the plugin opens a local app and keeps the loop visible: prompt-level evidence, candidate comparison, champion skill, failed mutations, history, and the next thing the evidence says to try. Dead ends stay in the record so the system does not keep rediscovering the same bad idea.

## Media

- Hero YouTube video: https://www.youtube.com/watch?v=Ymt1FzdO-IQ
- Primary image: https://www.justinwetch.com/skillrsi/hero.png
- Gallery image: https://www.justinwetch.com/skillrsi/hero.png
- Gallery image: https://www.justinwetch.com/skillrsi/codex-invoke.png
- Gallery image: https://www.justinwetch.com/skillrsi/codex-setup-sidebar.png
- Gallery image: https://www.justinwetch.com/skillrsi/codex-result-sidebar-dark.png
- Gallery image: https://www.justinwetch.com/skillrsi/running-live.png
- Gallery image: https://www.justinwetch.com/skillrsi/evidence-summary.png
- Gallery image: https://www.justinwetch.com/skillrsi/evidence-detailed-breakdown.png
- Gallery image: https://www.justinwetch.com/skillrsi/evidence-prompts.png
- Gallery image: https://www.justinwetch.com/skillrsi/evidence-prompt-expanded.png
- Gallery image: https://www.justinwetch.com/skillrsi/candidate-compare.png
- Gallery image: https://www.justinwetch.com/skillrsi/champion-skill.png
- Gallery image: https://www.justinwetch.com/skillrsi/history.png
- Gallery image: https://www.justinwetch.com/skillrsi/next-loop-plan.png
- Gallery image: https://www.justinwetch.com/skillrsi/diagram-core-loop.png
- Gallery image: https://www.justinwetch.com/skillrsi/diagram-ontology-map.png
- Gallery image: https://www.justinwetch.com/skillrsi/diagram-experiment-plan.png
- Gallery image: https://www.justinwetch.com/skillrsi/diagram-history-memory.png
- Gallery image: https://www.justinwetch.com/skillrsi/thumbnail.jpg

## Facts

- **Year:** 2026
- **Published:** 2026-06-02
- **Last updated:** 2026-06-17
- **Category:** AI, Open Source, Developer Tools
- **Tags:** AI, Agents, Skills, Evals, Codex, OpenAI, Open Source, Dev Tools, Automation, R&D
- **Canonical URL:** https://www.justinwetch.com/projects/skill-rsi/

## Links

- [GitHub repo](https://github.com/justinwetch/Skill-RSI) - source
- [Read the blog post](https://www.justinwetch.com/blog/skillrsi) - essay
- [Watch the intro video](https://youtu.be/Ymt1FzdO-IQ?si=qsA5WGTq0iBxc2Fj) - video

## Tags

- AI
- Agents
- Skills
- Evals
- Codex
- OpenAI
- Open Source
- Dev Tools
- Automation
- R&D