Caveman

Caveman is a skill/plugin that makes an AI coding agent answer in compressed "caveman-speak": filler and hedging are dropped, sentences become terse fragments, but code, commands, and error text are kept byte-for-byte exact. The point is to shrink what the agent says, not what it knows — the same technical answer in roughly a third of the words. It works with Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other agents, and preserves the user's own language (it compresses style rather than translating). Four compression levels are available — lite, full (default), ultra, and wenyan, which renders answers in classical Chinese for maximum meaning per token.

Beyond the core style skill it bundles related commands: terse Conventional Commit messages, one-line PR review comments, a session token-savings stats readout, a memory-file compressor that rewrites files like CLAUDE.md into caveman-speak (cutting their input-token cost in every future session), an MCP middleware that compresses tool descriptions of wrapped MCP servers, and compressed subagents for investigation, building, and review.

The project's own benchmarks, committed to the repo and reproducible, report an average 65% reduction in output tokens across ten tasks (range 22–87%). The README is unusually candid about the limits: only output tokens shrink, the skill itself adds roughly 1–1.5k input tokens per turn, whole-session savings run smaller than the headline number, and on already-terse workloads the net effect can be negative — the authors frame readability and speed as the primary win. The tool makes no network calls after install and collects no telemetry.