JuliusBrussee

caveman

AI🌱 NEW PROJECT BOOST#LLM #Claude #Prompt Engineering#Token Optimization#Agentic Workflow

// summary

Caveman is a specialized plugin for AI agents that significantly reduces output token usage by enforcing a concise, telegraphic communication style. It maintains full technical accuracy while cutting approximately 75% of output tokens and 46% of input tokens through its compression tools. The project supports various agents including Claude Code, Cursor, and Gemini, offering multiple intensity levels and specialized modes like 文言文.

// technical analysis

Caveman is a specialized plugin and skill set designed to optimize LLM interactions by enforcing a highly terse, 'caveman-speak' communication style that significantly reduces token consumption. By stripping away filler, pleasantries, and redundant prose while maintaining technical accuracy, the project achieves an average 75% reduction in output tokens and a 46% reduction in input tokens via its compression tool. The architecture is built on the philosophy that brevity improves readability and speed without sacrificing the quality of technical output, supported by empirical benchmarks and research on brevity constraints.

// key highlights

Reduces output token usage by approximately 75% by removing fluff while preserving technical substance.

Includes a compression tool that cuts input tokens by ~46% by rewriting memory files into a concise format.

Offers multiple intensity levels, including a unique 'Wenyan' mode for extreme classical Chinese compression.

Provides specialized commands like /caveman-commit and /caveman-review for efficient version control and code analysis.

Supports a wide range of AI agents including Claude Code, Cursor, Windsurf, Gemini CLI, and GitHub Copilot.

Improves response speed and readability by eliminating walls of text in favor of telegraphic, high-signal communication.

// use cases

Drastically reduce LLM API costs and latency by compressing agent responses.

Generate terse, professional commit messages and one-line code reviews.

Compress long-form memory files like CLAUDE.md to save input tokens across sessions.

// getting started

To begin, identify your preferred AI agent from the provided installation table and run the corresponding command, such as 'claude plugin install caveman@caveman' for Claude Code or 'npx skills add JuliusBrussee/caveman' for other supported agents. Once installed, you can trigger the mode using commands like '/caveman' or by simply asking the agent to 'talk like caveman'. For agents without native hook support, you can ensure always-on functionality by pasting the provided system prompt snippet into your agent's rules or configuration file.