caveman
Cut ~75% of output tokens by talking like caveman. Full technical accuracy. Lite, full, ultra, and 文言文 intensity levels. Works on Claude Code, Cursor, Copilot, Gemini CLI, and 40+ agents.
SKILL.mdagent/skills/caveman/SKILL.md
GitHubCaveman
Community skill — source and full install at github.com/JuliusBrussee/caveman
Ultra-compressed communication mode. Cuts token usage ~65–75% by having the agent talk like caveman — dropping articles, filler, and pleasantries — while keeping every bit of technical accuracy.
Install
# Any agent (auto-detect)
npx skills add JuliusBrussee/caveman
# Claude Code
claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
# Cursor
npx skills add JuliusBrussee/caveman -a cursor
# Windsurf
npx skills add JuliusBrussee/caveman -a windsurf
# Gemini CLI
gemini extensions install https://github.com/JuliusBrussee/caveman
Usage
/caveman — activate (default: full)
/caveman lite — drop filler, keep grammar
/caveman full — classic caveman (fragments OK, drop articles)
/caveman ultra — maximum compression, arrows for causality
/caveman wenyan — classical Chinese 文言文 mode
stop caveman — back to normal
Companion Skills
caveman-commit— terse commit messages, ≤50 char subjectcaveman-review— one-line PR comments/caveman:compress— compress CLAUDE.md / memory files (~46% input token savings)
Good to Know
Advanced reference for Caveman. Background, edge cases, and patterns worth understanding.
Output tokens vs. input tokens
Caveman cuts output tokens — what the model writes back to you. It does not affect input tokens (what you send or what gets loaded from CLAUDE.md, memory files, etc.).
For input token savings, use caveman-compress: it rewrites your CLAUDE.md and memory files into caveman-speak so Claude reads less at session start (~46% savings per file). Both together give you the full picture.
| What it saves | Tool |
|---|---|
| Output tokens (responses) | caveman |
| Input tokens (CLAUDE.md, memory) | caveman-compress |
Technical accuracy is not compressed
The skill explicitly preserves:
- Technical terms and proper nouns
- Error messages (quoted exactly)
- Code blocks, file paths, URLs
- Version numbers and flags
- Warnings and caveats
Only prose fluff is removed. A March 2026 paper found that brevity constraints on LLMs can improve accuracy by 26 percentage points on some benchmarks — less rambling, more precise answer.
Auto-clarity for dangerous operations
Caveman automatically suspends compression for:
- Destructive operations —
DROP TABLE,rm -rf, irreversible actions - Security warnings — anything that could cause data loss or exposure
- Multi-step sequences — where fragment ordering risks misread
After the clear part is done, caveman mode resumes automatically. You don't need to toggle it.
Intensity levels
| Level | Command | What it does | Best for |
|---|---|---|---|
| lite | /caveman lite |
Drop filler, keep grammar | Readability-sensitive outputs |
| full | /caveman full |
Fragments OK, drop articles | Default daily use |
| ultra | /caveman ultra |
Max compression, → for causality |
Long sessions, heavy context |
| wenyan | /caveman wenyan |
Classical Chinese 文言文 | Extreme compression experiment |
Start with full. Drop to lite if teammates read your AI output. Use ultra only when context is tight.
Works across all agents
Caveman is agent-agnostic. It works via system prompt injection, so any agent that accepts a system prompt or rules file is compatible:
Claude Code · Cursor · Windsurf · Gemini CLI · GitHub Copilot · Aider · Continue · Cline · and 40+ more
Claude Code: auto-activate every session
Add this to your CLAUDE.md so caveman activates at session start without a manual /caveman command:
<!-- CLAUDE.md -->
/caveman full
Or for project-specific intensity:
/caveman ultra
This fires at the top of every session. Pair with caveman-compress on the same file to cut both input and output costs.
The benchmark numbers are real
The 65–75% output token reduction is measured, not estimated. Numbers come from real sessions across coding, debugging, and explanation tasks. Variation depends on:
- Task type — explanations compress more than code generation
- Intensity level —
ultrasaves more thanlite - Model verbosity baseline — Claude 3.x tends to be more verbose than GPT-4o, so savings are higher
The 26% accuracy improvement cited above is from external research on brevity constraints, not a caveman-specific benchmark. But the directional finding holds: tighter output ≠ worse answers.