caveman

caveman-translate

Caveman mode with EN↔ZH translation. Type in English, Claude thinks in Mandarin Chinese, you see English.

Live Benchmark

https://caveman-benchmark.surge.sh

Token comparison across 10 coding prompts using Claude MiniMax M2.7 API.

Results

Mode	Avg Input	Avg Output	Avg Total	vs Baseline
Baseline	31	837	868	—
caveman-full	183	369	552	-36%
caveman-translate	267	475	743	-14%

caveman-full cuts output tokens ~56% (837 → 369) with 36% total token savings.

caveman-translate adds ~85 input token overhead from wenyan rules + translation context but doesn't compress as aggressively in English output.

How It Works

1. You type prompt in English
2. UserPromptSubmit hook translates EN → ZH via Google Translate API
3. Claude receives Mandarin prompt, responds in Mandarin (wenyan mode)
4. caveman-translate MCP middleware translates ZH → EN transparently
5. You see English output

Setup

1. Google Translate API Key

Get a key from Google Cloud Console:

• Create project → Enable "Cloud Translation API" → Create API Key
• Free tier: 500K chars/month

2. Configure Claude Code

Add to your settings.json:

{
  "env": {
    "GOOGLE_TRANSLATE_API_KEY": "your-key-here",
    "CAVEMAN_TRANSLATE": "1"
  },
  "plugins": {
    "installed": [
      { "path": "/path/to/caveman-translate" }
    ]
  }
}

3. Activate

/caveman wenyan-full

What Gets Translated

• Prompts: EN → ZH before Claude sees them
• Responses: ZH → EN before you see them
• Code blocks, URLs, file paths: not translated

Cavecrew Subagents

Caveman ships with three subagents for delegated work:

• cavecrew-investigator — read-only code locator. Finds symbols, paths, usages.
• cavecrew-builder — surgical 1-2 file editor. Refuses 3+ file scope.
• cavecrew-reviewer — diff/file reviewer. One-line findings with severity.

File Structure

caveman/
├── skills/caveman/           # Wenyan mode skill
├── src/
│   ├── hooks/                # Claude Code hooks (activate, config, mode-tracker)
│   └── mcp-servers/
│       └── caveman-translate/  # ZH→EN MCP middleware
└── benchmarks/
    ├── index.html            # Live benchmark dashboard
    └── results.json          # Raw token data

Benchmarks

Run your own benchmarks:

cd benchmarks
node run.js

Requires ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL in environment.

License

MIT