The gist

Caveman is an open-source skill and plugin developed by Julius Brussee that optimizes AI agent communication. It condenses LLM output into a terse, "caveman-speak" style, significantly reducing token usage while retaining technical accuracy. The tool addresses the problem of verbose AI responses, making them faster, cheaper, and easier to read for developers.

What it does

Condenses AI agent output by up to 87% for reduced token usage across various prompts.
Maintains full technical accuracy while simplifying language and removing conversational fluff.
Provides customizable terseness levels, including Lite, Full, Ultra, and Classical Chinese modes.
Generates terse commit messages following Conventional Commits standards for clarity.
Creates concise, one-line pull request review comments.
Compresses input memory files, reducing LLM input tokens by an average of 46%.

How it works

Caveman is installed as a plugin or skill into various AI agents like Claude Code, Gemini CLI, Cursor, and Copilot. Users trigger its modes or skills via slash commands or special syntax. It processes LLM output to strip filler words and simplifies grammar, delivering concise, technically accurate responses. It also offers an input compression tool. The project is open-source and free to use under an MIT license.

Best for

Developers interacting with AI coding assistants will find Caveman ideal for optimizing their workflow. It drastically cuts down on verbose AI responses and input context, making interactions faster, cheaper, and more direct for technical tasks like code review or debugging.

Watch out for

While highly effective for output compression, Caveman primarily impacts output tokens; thinking and reasoning tokens remain untouched. Auto-activation behavior varies across agents, sometimes requiring manual configuration for always-on functionality.