JuliusBrussee/caveman

⭐ 57,276  ·  Python  ·  GitHub Repo

🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman

ai anthropic caveman claude claude-code llm meme prompt-engineering

1-Sentence Summary

Caveman slashes AI token costs 65%+ by forcing agents to speak like cavemen—genius, practical, and hilarious.

🔥 Key Capabilities & USP

• Massive Token Compression (65-75% Output, ~46% Input): Strips filler words, articles, and fluff while preserving full technical accuracy. Solves the pain point of runaway token costs and slow response times in AI coding assistants.
• Universal Agent Support (30+ Agents): Single install command works across Claude Code, Gemini CLI, Codex, Cursor, Windsurf, Cline, Copilot, and more. Eliminates the headache of configuring token-saving settings per tool.
• Multiple Intensity Levels: Choose from Lite, Full, Ultra, and 文言文 (Classical Chinese) modes. Lets you dial compression from "polite brevity" to "caveman grunts" based on context.
• Caveman Skills Suite: Specialized commands like /caveman-commit (terse commit messages), /caveman-review (one-line code reviews), and /caveman-stats (lifetime token savings). Turns token optimization into a fun, gamified workflow.
• MCP Middleware (caveman-shrink): Proxy that compresses input tokens by ~46% every session, compounding savings on both sides of the conversation.

USP: Caveman transforms a viral meme about LLM verbosity into a production-grade, cross-agent token optimization tool that's both practical and delightful to use.

Technical Architecture

Component	Technology	Purpose
Installation	Shell script (install.sh) + PowerShell (install.ps1)	Auto-detects installed agents, runs native installs for each
Plugin System	npx skills framework	Cross-agent compatibility for 30+ AI coding tools
Native Integrations	Claude Code plugin system, Gemini CLI extensions	Direct support for major agent ecosystems
MCP Middleware	caveman-shrink proxy	Compresses input tokens (~46%) at the transport layer
Per-Repo Rules	.cursor/rules/caveman.mdc, .windsurf/rules/caveman.md, .clinerules/caveman.md, .github/copilot-instructions.md, AGENTS.md	Automatic activation per repository
Plugin Modes	Lite, Full, Ultra, 文言文	Configurable compression intensity via prompt engineering

Quick Start Guide

Installation (macOS/Linux/WSL/Git Bash):

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

Installation (Windows PowerShell):

irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

Full installation (plugin + hooks + statusline + MCP shrink + per-repo rules):

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash -s -- --all

Minimal installation (plugin only):

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash -s -- --minimal

Dry run (preview only):

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash -s -- --dry-run

Pros, Cons & Use Cases

Pros

• Massive cost savings: 65-75% output token reduction directly lowers API bills for heavy AI agent users.
• 3x faster responses: Less verbose output means faster completion times.
• Full technical accuracy preserved: Strips fluff, not substance. Code, commands, and technical details remain intact.
• One install, 30+ agents: No need to configure each tool separately.
• Fun factor: Code reviews become comedy. Lowers friction for team adoption.

Cons

• Reduced readability for complex explanations: Non-technical stakeholders or junior devs may struggle with ultra-terse responses.
• Ultra mode can be too telegraphic: May require switching to Lite mode for nuanced discussions.
• Session-based activation: Requires re-activation per session unless using auto-start features.
• 文言文 mode niche: Classical Chinese mode is only useful for users fluent in that language.

Who should NOT use this?

• Teams with non-technical stakeholders who rely on AI-generated explanations for decision-making.
• Junior developers who benefit from verbose, pedagogical responses from AI agents.
• Documentation-heavy workflows where AI generates prose for end-user consumption.
• Users on extremely tight budgets who cannot afford even the minimal token overhead of the plugin itself.

Ideal Use Cases

• Solo developers and power users who want faster, cheaper AI coding assistance.
• CI/CD pipelines where AI agents generate commit messages, code reviews, and changelogs.
• Teams with high AI agent usage looking to slash monthly API costs by 50-70%.
• Hackathons and rapid prototyping where speed matters more than conversational polish.
• Terminal-first workflows where users prefer terse, command-like responses.

Community & Activity

Caveman has exploded onto the scene with 57,276 stars—a clear signal that the developer community has been waiting for a practical, fun solution to AI token bloat. The project is actively maintained (last updated May 2026) and has spawned a rich ecosystem of intensity modes and agent integrations. The viral "why use many token when few token do trick" tagline perfectly captures the zeitgeist of developer frustration with verbose AI responses. With this level of momentum, Caveman is well on its way to becoming the de facto standard for token optimization in AI coding assistants.