Caveman Plugin Cuts AI Coding Agent Verbosity Without Reducing Reasoning

A new tool called Caveman is a skill/plugin designed to make AI coding agents communicate in shorter, direct language while keeping code, commands, and error outputs fully intact. The project targets the execution phase of development, where verbose agent responses — restating tasks, explaining obvious steps, and summarizing repeatedly — slow down workflows without adding value. Caveman claims to reduce output tokens by around 65%, though the developers note this figure varies by agent, task, and prompt, and should not be treated as a universal benchmark. It is compatible with several popular agent environments including Claude Code, Codex, Gemini, Cursor, and GitHub Copilot. The core idea is to shrink what the agent says, not what it thinks, keeping technical findings, uncertainties, and verification steps intact while stripping out conversational filler.