Gemini Evolve — Development skill for Claude Code
Self-evolve Gemini CLI instructions, commands, and skills via the gemini CLI itself — GA + GEPA/DSPy, with hard gates before apply.
How to install Gemini Evolve
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open LichAmnesia/gemini-evolve and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Gemini Evolve does
Self-evolve Gemini CLI instructions, commands, and skills via the gemini CLI itself — GA + GEPA/DSPy, with hard gates before apply.
Alternatives in Development
- Understand Anything — Graphs that teach graphs that impress 80.7k ★
- Career-Ops For Codex — Read CLAUDE.md for all project instructions, routing, and behavioral rules 32.5k ★
- Vibe Kanban — Get 10X more out of Claude Code, Gemini CLI, Codex, Amp and other coding agents 23.5k ★
README
gemini-evolve
English | [简体中文](README.zh.md)
**Stop hand-tuning `GEMINI.md`. Let Gemini tune it — with holdout tests and hard gates, so regressions never ship.**
If you use the [Gemini CLI](https://github.com/google-gemini/gemini-cli), you already have a `GEMINI.md` somewhere. If it's been growing for a while, chances are some of it has gone vague, duplicated, or stale. `gemini-evolve` runs a small optimization loop that **mutates your instructions, grades each variant by actually running Gemini CLI on real tasks, and only writes back when a holdout set (examples it never trained on) says the new version is genuinely better.**
Before → after (illustrative example, not a benchmark)
Your baseline `~/.gemini/GEMINI.md`:
# My Gemini Instructions
- try to be concise
- generally avoid long explanations
- if possible, use bullet points
- prefer to keep answers short
After one `gemini-evolve evolve --apply` run, the `sharpen_constraints` + `condense` mutations typically produce something like:
# My Gemini Instructions
## Response Style
- Answer in <=3 sentences by default; use bullets for lists of 3+ items.
- No preamble ("Sure!", "Great question!"). Start with the answer.
- Expand only when the user asks "explain" / "why" / "walk me through".
Holdout score went up, size cap not exceeded, so `--apply` wrote the new file and kept `GEMINI.md..bak` as your escape hatch. If the holdout score had *not* improved by at least 2%, the loop would refuse to apply — no sneaky regressions.
Why you might want this
- You've edited
GEMINI.mdby hand one too many times. - You suspect it's bloated but don't want to break what works.
- You want measured improvements, not vibes.
Prerequisites
| Requirement | How to check | Notes |
|---|---|---|
| Python 3.11+ | python3 --version |
3.12 tested too |
gemini CLI installed + logged in |
gemini --version then gemini -p "hi" -o json |
Install: [googl |
Related Skills
Hermes Agent Self Evolution
⚒ Evolutionary self-improvement for Hermes Agent — optimize skills, prompts, and code using DSPy + GEPA
Claude SEO — Multi-Platform Agent Instructions
For Cursor, Cursor Cloud Agents, Google Antigravity, and Gemini CLI. Claude Code users: see CLAUDE.md instead.
Enforcement Contradiction Rules
Always-on. When the enforcement system itself produces contradictions — gates blocking their own prerequisites
Otherside CLI
Totally Free Claude Code inspired terminal coding agent. Multi-provider: Anthropic, Codex, Gemini, Kimi, DeepS
PrAImate
A terminal launcher for agent CLIs — Claude Code, Codex CLI, OpenCode, Gemini CLI, DeepSeek-TUI — that pairs e
Omnicode
Self-hosted MCP hub: register an MCP server once from your phone and use it in Claude Code, Codex CLI and Gemi
Related Agents
Constraint Check
Internal affairs auditor — checks whether HARD-GATEs and constraints were actually followed during workflow. U
Shux Tui QA
General hard-gate QA subagent for shux TUI, terminal UX, CLI, pane automation, plugin, attach, mouse/input, an
Prompt Polisher
Analyzes AI prompt files (.claude/agents/.md, .claude/skills//.md, .agents/skills//.md, CLAUDE.md, AGENTS.md,