I built a self-evolving AI skill engine as a solo dev using 4 AI models

What is Cambrian?

An AI skill engine where skills get better the more you use them.

Every AI tool runs the same prompts forever. Cambrian is different:

• Give feedback → LLM mutates the skill
• 3-trial benchmark → LLM Judge blind-scores A/B
• Better version survives, worse one dies
• After 10 rounds, the skill is personalized to YOUR workflow

How I built it

Solo Korean developer. 4 AI models in relay:

• Claude Opus → architecture & critical thinking (40 rounds of self-critique)
• GPT → code generation
• Claude Sonnet → implementation & debugging
• Gemini → review

The project itself was evolved through its own philosophy:
feedback → improve → test → repeat.

Key features

• 3-channel feedback: auto (failure analysis), critic (proactive review), manual
• LLM Judge: blind A/B scoring (0-10) with reasoning
• Competitive execution: same domain, best skill wins
• Multi-provider: Claude, GPT, Gemini — not locked to one platform
• Skill decay: 30 days unused → dormant, 90 days → fossil
• Export/import: portable .cambrian packages

Numbers

• 180 tests, 0 failures
• 14 seed skills
• 16 CLI commands
• 3 LLM providers

Try it

pip install git+https://github.com/Ait426/cambrian.git
cambrian init --dir ./my_project
cambrian skills

GitHub: https://github.com/Ait426/cambrian

What would you break first?