GEPA optimize-anything plugin
A GEPA plugin for Claude Code: the agent builds an evaluator and searches for a better prompt, code or config by score.
- 7,415 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The evaluator runs your code and some engines start a Claude Code subprocess and spend tokens.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/gepa-optimize-anythingDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The evaluator runs your code and some engines start a Claude Code subprocess and spend tokens.
This is third-party code. Review the repository files before installing.
What it does
The `gepa-optimize-anything` plugin teaches an agent to drive the `optimize_anything` API. A seed text artifact (prompt, program, config, SQL, regex, agent scaffold) improves in a loop: the optimizer drafts an edit, the evaluator scores it and returns feedback. The user defines the evaluator. It can be a metric (test pass rate, latency, cost) or an LLM judge for subjective tasks. Search runs on the `gepa` engine (reflective evolutionary search, the default), `autoresearch` or `meta_harness`; a naive `best_of_n` serves as a baseline. The skill ships with the repository, and the Python `gepa` package is part of the same project.
Who it is for. Prompt and agent developers who want reproducible scoring and automated search instead of hand edits.
Good fit when
- A system prompt fails on part of the test cases and needs automated search.
- Several optimizers must be compared on one task and one evaluator.
- A config or SQL query is scored by latency or cost.
- An agent scaffold is tuned against a set of check tasks.
Not a fit when
- The artifact has no quality signal at all: no metric and no judge.
- Evaluator runs are expensive and there is no budget for hundreds of calls.
- A one-off manual text edit is enough.
Example request
Improve the prompt in `prompt.txt` with optimize_anything: write an evaluator over 20 test examples, run the gepa engine and show how the score changed.Limitations
Result quality is capped by the evaluator and its feedback: a poor score steers search the wrong way. Every iteration calls models and the evaluator, so cost grows with the number of runs. The `autoresearch` and `meta_harness` engines start a Claude Code subprocess and the evaluator runs your code; use an isolated working directory. The project changes quickly and the plugin is at version 0.1.0. Check optimization results on a held-out set: tuning against one evaluator may degrade behavior outside the test set.
How to disable. Run `/plugin uninstall gepa-optimize-anything@gepa` and optionally remove the `gepa` marketplace. Remove the `gepa` Python package from your environment separately.
FAQ
How do I install the plugin?
Run `/plugin marketplace add gepa-ai/gepa`, then `/plugin install gepa-optimize-anything@gepa`. In a clone of the repository Claude Code discovers the skill on its own.
Which agents work with the skill?
The README names Claude Code and agents that read `.claude/skills/`: Cursor, VS Code with Copilot, Codex and Gemini CLI.
How is GEPA different from optimize_anything?
`optimize_anything` is the general text-optimization API. GEPA is the default optimizer behind it and also the name of the Python package.
What does a run cost?
Every iteration calls models and the evaluator, so cost depends on the run budget. Start with few iterations.
Can I run it without a metric?
Yes, an LLM judge can serve as the evaluator for subjective tasks. Score quality still determines the result.
Discussion
No questions or reviews yet.
Related
PluginsAgent self-improvement framework: optimizes prompts, memory, skills or harness code with method and evaluation set separately.
cap-evolve edits an agent's prompts, tools and skills from failed traces and keeps an edit only if it wins on a held-out split
A plugin that cuts token use in long sessions: prunes stale tool output, picks a subagent model and narrows the skill list.
A hook plugin that adds context to a Claude Code prompt and asks one to six grounded questions when a request is vague.
A Claude Code plugin with hooks and 26 skills: measures what the agent knows, blocks edits until code is read, keeps memory
AI Builder Club Skills
AI Builder Club — Skills
A Claude Code plugin marketplace to make a repo agent-ready and run compounding agent loops
README in short
The project presents GEPA as a reflective optimizer for text artifacts and describes installing the skill from the repository marketplace.