GitHub - JuliusBrussee/caveman - Claude Code skill that cuts 65% of tokens by talking like caveman

Abstraction: Prompt skill compressing AI agent output 65% via caveman-style language

Key points:

• Installs as a skill/plugin for Claude Code, Gemini CLI, Cursor, Windsurf, Cline, Codex, and 30+ other agents; cuts output tokens by average 65% (range 22–87%) while preserving 100% technical accuracy
• Five compression levels: lite, full (default), ultra, wenyan (classical Chinese for maximum token density)
• Companion tools include /caveman-compress for shrinking memory files like CLAUDE.md (~46% input token savings per session), /caveman-stats for real-time savings tracking, and caveman-shrink MCP middleware
• A March 2026 paper found that brevity constraints improved accuracy by ~26 points on some benchmarks for large models
• Savings are output-token-only; the skill adds ~1–1.5k input tokens per turn, so whole-session savings are smaller than headline number
• No telemetry, no backend; caveman is a local prompt + hooks, MIT licensed

Connections: Claude Code · Julius Brussee · Token Optimization · Prompt Compression · Agentic Coding

Source: https://github.com/juliusbrussee/caveman