Skip to content

README.md

Latest commit

 

History

History
107 lines (72 loc) · 2.81 KB

README.md

File metadata and controls

107 lines (72 loc) · 2.81 KB

claude-evolve 🧠

Your CLAUDE.md learns automatically from every correction you make.

You correct Claude. It never learns. Next session — same mistake.

claude-evolve fixes that.

Every time you correct Claude Code, it silently extracts the rule and writes it to your CLAUDE.md. Permanently. Zero manual work.

What it does

You:    "Don't use console.log, use our logger utility"
Claude: [continues working]

claude-evolve: [background, silent]
  → Detected correction
  → Extracted rule: "Use logger utility instead of console.log"
  → Written to CLAUDE.md ✓

Next session. Claude reads CLAUDE.md. Never makes that mistake again.

Install

# Inside Claude Code
/plugin marketplace add yourusername/claude-evolve
/plugin install claude-evolve

Or manual:

curl -fsSL https://raw.githubusercontent.com/yourusername/claude-evolve/main/install.sh | bash

Requires: ANTHROPIC_API_KEY in your environment.

How it works

  1. UserPromptSubmit hook fires on every message you send
  2. Pattern matching detects corrections in < 1ms — zero latency
  3. Confirmed corrections spawn a background process (async — never blocks you)
  4. Background process calls Claude Haiku API (~$0.001) to extract the rule
  5. Rule is deduped and appended to your CLAUDE.md
  6. Next Claude Code session — the rule is in context from the start

What gets learned

Your correction Rule added to CLAUDE.md
"Don't use any, use proper types" "Never use TypeScript any — always define proper types"
"Stop adding console.logs" "Do not add console.log statements unless explicitly requested"
"Use our API client, not fetch directly" "Use the internal API client (src/lib/api.ts), not raw fetch"
"Always run tests after changes" "Run the test suite after every significant code change"
"Don't commit directly to main" "Never commit directly to main — always use feature branches"

Privacy

  • Everything runs locally
  • The only external call is to Anthropic API (same as Claude Code itself)
  • Rules stay in your CLAUDE.md — your codebase, your control
  • Add <private> tags to exclude sensitive prompts from processing

Cost

~$0.001 per correction extracted. Haiku model.

If you make 50 corrections a month — that's $0.05/month.

Configuration

// ~/.claude-evolve/config.json
{
  "enabled": true,
  "minConfidence": 60,
  "maxRulesPerSession": 10,
  "excludePatterns": ["password", "secret", "token"]
}

Star History

Star History Chart

Built by an engineer with 25 years of production scars. Because CLAUDE.md should learn from you — not the other way around.