Caveman cuts ~75% of Claude's output tokens without losing technical accuracy. One-line install for Claude Code, Cursor, Windsurf, Copilot, and more. Four grunt levels, terse commits, one-line PR reviews, and input compression built in. 24.9K stars.
Julius taught Claude to talk like a caveman. 24.9K stars later, it's the most useful meme in developer tooling.
The problem: LLMs are verbose by default. Phrases like "I'd be happy to help you with that" and "Let me summarize what I just did" contribute nothing — but burn tokens, slow responses, and push you into usage limits faster.
The solution: Caveman makes Claude skip the throat-clearing and go straight to the answer. Same fix. 75% less word. Brain still big.
What stands out: 🪨 ~75% output token reduction: Benchmark average 65%, range 22–87% across real coding tasks ⚡ ~3x faster responses: Less token to generate = speed go brrr 🎚️ Four intensity levels: Lite, Full, Ultra, and 文言文 (Classical Chinese) mode 📝 Caveman-commit: Terse commit messages, ≤50 char subject, why over what 🔍 Caveman-review: One-line PR comments: L42: 🔴 bug: user null. Add guard. 🗜️ Caveman-compress: Rewrites your CLAUDE.md into caveman-speak, cutting ~46% of input tokens every session 🔌 Works everywhere: Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and 40+ more 🆓 Free, MIT, one-line install
Before and after: 🗣️ Normal Claude (69 tokens): "The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle..." 🪨 Caveman Claude (19 tokens): "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."
Note: works best for coding tasks. Nuanced responses still need full Claude, and the system prompt loads as input tokens, so net savings vary per use case.
A March 2026 paper found brevity constraints improved accuracy by 26 percentage points on certain benchmarks. Verbose not always better.
Perfect for developers hitting usage limits and anyone who wants their AI agent to do the work and shut up about it.
Rohan Chaubey
Julius taught Claude to talk like a caveman. 24.9K stars later, it's the most useful meme in developer tooling.
The problem: LLMs are verbose by default. Phrases like "I'd be happy to help you with that" and "Let me summarize what I just did" contribute nothing — but burn tokens, slow responses, and push you into usage limits faster.
The solution: Caveman makes Claude skip the throat-clearing and go straight to the answer. Same fix. 75% less word. Brain still big.
What stands out:
🪨 ~75% output token reduction: Benchmark average 65%, range 22–87% across real coding tasks
⚡ ~3x faster responses: Less token to generate = speed go brrr
🎚️ Four intensity levels: Lite, Full, Ultra, and 文言文 (Classical Chinese) mode
📝 Caveman-commit: Terse commit messages, ≤50 char subject, why over what
🔍 Caveman-review: One-line PR comments: L42: 🔴 bug: user null. Add guard.
🗜️ Caveman-compress: Rewrites your CLAUDE.md into caveman-speak, cutting ~46% of input tokens every session
🔌 Works everywhere: Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and 40+ more
🆓 Free, MIT, one-line install
Before and after:
🗣️ Normal Claude (69 tokens): "The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle..."
🪨 Caveman Claude (19 tokens): "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."
Note: works best for coding tasks. Nuanced responses still need full Claude, and the system prompt loads as input tokens, so net savings vary per use case.
A March 2026 paper found brevity constraints improved accuracy by 26 percentage points on certain benchmarks. Verbose not always better.
Perfect for developers hitting usage limits and anyone who wants their AI agent to do the work and shut up about it.