What Is caveman?
caveman is a Claude Code skill created by Julius Brussee that instructs Claude to communicate like a caveman: short sentences, no filler, no pleasantries, no hedging. Technical content remains intact. Code blocks are unchanged. Error messages are quoted exactly. Only the verbal scaffolding around the facts is stripped away.
The result is an average 65% reduction in output tokens across real-world coding tasks, validated by a published benchmark across 10 task types. Key capabilities:
65% average token reduction
Benchmarked across explain, fix, refactor, architecture, and review tasks
Caveman-style minimal responses
Drops articles, filler words, and pleasantries; keeps all technical substance
Six intensity levels
Lite (removes hedging only) → full → ultra → classical Wenyan (maximum compression)
Auto-triggers
Activates on ""be brief"" or ""fewer tokens""; reverts automatically for security warnings or destructive operations
Compatible with Claude Code, Codex, Cursor, and 40+ agents
How It Works
caveman operates as a SKILL.md file that sets a communication persona. When active, Claude drops:
- Articles: a, an, the
- Filler words: just / really / basically / actually / simply
- Pleasantries: Sure / Certainly / Happy to help! / Great question! / Hedging language / It's worth noting that… / You might want to consider…
- Preamble and postamble: no restating the question, no summaries of what was done
What stays:
- All technical terms, exactly as written
- Code blocks, completely unchanged
- Error messages, quoted exactly
- Factual content, complete and accurate
Before and After
Normal Claude: I've successfully completed the refactoring of the authentication module. The changes include updating the token validation logic to handle edge cases more gracefully, adding appropriate error handling, and ensuring backwards compatibility with the existing API contracts.
Caveman Claude: Auth refactor done. Token validation updated. Error handling added. API-compatible.
Same information. 75% fewer tokens.
Token Savings Benchmark
65% Average Output Reduction: Based on benchmarks across 10-11 actual development tasks, caveman mode averaged 294 tokens per response, compared to 1,214 tokens in normal mode.
22%–87% Range: Savings vary between 22% and 87% depending on the specific prompt.
65% Total Reduction (Reported): Some users report saving up to 60% of their total daily token usage.
Input Token Compression (Caveman Compress): A companion tool, caveman-compress, specifically targets memory files (like CLAUDE.md), reducing their token count by roughly 40-46%.
How to Install caveman?
npx skills add juliusbrussee/caveman --skill caveman
Restart Claude Code. Caveman is active from message one, no activation phrase needed.
Or download from GitHub https://github.com/JuliusBrussee/caveman