Caveman is a JavaScript-based skill and plugin for Claude Code and 30+ other AI coding agents that reduces output token usage by compressing agent responses into terse, caveman-style speech while preserving technical accuracy. The core mechanism strips filler language, redundant explanations, and verbose phrasing from agent replies, keeping code, commands, error messages, and URLs byte-for-byte exact. According to benchmarks in the repository, this approach achieves a 65% reduction in output tokens on prose-heavy responses like explanations and debugging walkthroughs, though the savings drop to 8.5% on full agentic coding runs where most output consists of code and tool calls rather than narration.
The skill installs via a single command that automatically detects and configures every agent on a user's machine, supporting Node 18 and above. Once installed, users can toggle caveman mode on or off with commands like `/caveman` or by saying "talk like caveman," and the behavior persists for the session. The repository includes six compression levels ranging from lite to ultra, plus a specialized wenyan mode that renders responses in classical Chinese for additional token density. Each level can be switched at any time with `/caveman <level>`.
Beyond the core compression feature, Caveman provides several supplementary commands. The `/caveman-review` command produces one-line PR comments in a compressed format. The `/caveman-stats` command displays real session token usage, lifetime savings, and USD cost, with an option to generate a shareable tweet. The `/caveman-compress` command rewrites memory files like CLAUDE.md into caveman-speak, cutting approximately 46% of input tokens on every subsequent session while preserving code, URLs, and paths. The repository also includes caveman-shrink, an MCP middleware tool available on npm that wraps any MCP server and compresses its tool descriptions, plus cavecrew subagents for investigation, building, and review that operate at roughly 60% of vanilla token usage.
The benchmarks presented in the repository show measured token counts from the Claude API across ten chat-style prompts, with output reduction ranging from 22% to 87% and averaging 65%. These measurements are committed and reproducible in the benchmarks and evals directories. An independent evaluation by JetBrains in July 2026 tested the skill against 86 tasks from SkillsBench using Claude Code on claude-sonnet-5, confirming the 65% prose reduction while documenting an 8.5% savings on full agentic coding runs. The repository explicitly acknowledges that the skill only compresses output tokens, leaving input and reasoning tokens untouched, and that the skill itself adds 1 to 1.5k input tokens per turn, making whole-session savings smaller than the output-only number suggests.
The repository is classified under AI, Anthropic, Claude, Claude Code, LLM, prompt engineering, and meme topics. It maintains active development with a homepage at caveman.so and includes comprehensive installation documentation in INSTALL.md. The approach is language-agnostic, preserving the user's original language while compressing style, with the exception of wenyan mode which intentionally renders in classical Chinese for maximum token efficiency.
