Agent Skill · Claude Code plugin · MIT
Recursive self-improvement for Claude Code skills
rsi-skill is an open-source recursive self-improvement (RSI) loop for Claude Code, Codex and Agent Skills. Point it at any SKILL.md and a goal: it writes fresh eval cases every iteration, runs the skill headless, grades the runs with deterministic and LLM-as-a-judge checks, and keeps a change only when it wins on cases nobody has seen.
Install
Claude Code
/plugin marketplace add zsu/agent-plugins
/plugin install rsi-skill@zsu
Codex: codex plugin marketplace add zsu/agent-plugins, then install
rsi-skill from the Plugins directory. Test runs use the claude CLI headless, so Claude Code
must be installed and signed in.
Run it
/rsi-skill:rsi-skill run against <path-to-skill> until no major problems, max 4 iterations, $60
Goals can be "no major problems", "pass rate ≥ 90%" or a focus such as "unattended mode". Limits can be iterations, time, tokens or cost.
The loop
- Profiles the skill and turns your goal into a measurable
goal.json. - Writes new eval cases in every iteration, so the skill is never tuned to a fixed benchmark.
- Runs the skill headless (
claude -p) in a sandbox for each case. - Grades every run with deterministic checks first, then an evidence-required LLM grader.
- Lets an improver edit anything in the skill: SKILL.md, scripts, references, assets and tests.
- Promotes the challenger only when it beats the champion on the next iteration's unseen cases.
- Stops on goal, budget, plateau, or problems that are stuck or external.
The best version comes back as uncommitted edits to review with git diff. Its
evals.json, grading.json and benchmark.json use the same field names as
Anthropic's skill-creator.