Claude Evolve

claude-evolve

Claude Code watches what you do. This watches what Claude Code does, and makes it better at working with you.

Most "memory" tools for AI assistants just store and retrieve. claude-evolve **evolves** — it spots your mistakes before you correct them, learns patterns you never explicitly teach, prunes rules that stop helping, and gets sharper every session. After enough sessions, it graduates from rules to full **Claude Code skills** that capture how you think.

See it in action (2 minutes to set up, 10 minutes to see)

Don't want to read docs? Just pick a profession and watch the system learn:

git clone https://github.com/jack60810/claude-evolve.git
cd claude-evolve
node demo.js analyst          # or ios-engineer, game-developer, ml-ops, random...

The demo:

1. Asks Claude (sonnet) to generate realistic deep work sessions for that profession
2. Runs each session through the learning pipeline (just like a real user would)
3. Observes patterns, scores rules, promotes mature ones to a skill
4. Prints the final CLAUDE.md and .claude/skills/auto-*.md for you to read

You'll see, in real output, the skill the system learned — the user's **thinking model**, not just their steps. Different profession = different skill. No hardcoded templates.

Two example outputs (not for programmers)

To show this isn't just a coding tool — here are two skills the system produced, each from 3 simulated sessions of someone doing their real job. No code involved, no hardcoded templates, no pre-written rules.

Example 1: Kindergarten Principal — admin work (scheduling, classroom observations, budgets)

---
name: auto-kindergarten-ops-discipline
description: >
  This user thinks in terms of "verify before commit" — never finalize anything
  without cross-checking all dependencies first. They prioritize correctness and
  accountability over speed, and expect proactive conflict detection, not reactive fixes.