JuliusBrussee/caveman

A skill for Claude Code and 30+ other agents that makes the model answer in terse "caveman-speak," cutting output tokens while keeping code and commands exact.

What it is

Caveman is a skill/plugin that tells an AI coding agent to drop filler and reply in tight fragments — "caveman-speak" — while leaving code, commands, and error text byte-for-byte unchanged. The README reports about 65% fewer output tokens on average across its benchmark prompts, on the premise that it shrinks what the agent says, not what it knows. It is aimed at developers who find agent replies needlessly verbose and want shorter, faster-to-read answers on every turn.

Key features

Compresses reply style, not substance, and explicitly preserves code, commands, and errors verbatim; it keeps the user's language rather than translating.
Six intensity levels switchable with /caveman <level>: lite, full (default), ultra, and a wenyan mode that renders in classical Chinese for maximum meaning per token.
Extra commands: /caveman-commit (Conventional Commit messages), /caveman-review (one-line PR comments), /caveman-stats (session token usage and lifetime savings), and /caveman-compress (rewrite a memory file like CLAUDE.md into caveman-speak to cut input tokens on future sessions).
A caveman-shrink MCP middleware that wraps any MCP server and compresses its tool descriptions, plus cavecrew-* subagents (investigator, builder, reviewer).
A single installer that detects the AI coding agents on your machine and installs for each; on Claude Code, Codex, and Gemini a hook turns it on from the first message.
No telemetry, analytics, accounts, or backend; after install the skill makes zero network calls.

Tech stack

JavaScript; the installer package is caveman-installer (version 0.1.0), MIT licensed, requiring Node >=18.
Installation via a curl | bash / PowerShell script, a Node bin (caveman), or per-agent paths (Claude Code plugin, Gemini CLI extension, npx skills add).
Ships skills, commands, agents, and plugin bundles; local hook scripts (shell + PowerShell) drive session activation and the statusline; benchmarks/evals use Python (benchmarks/run.py, evals/).
The caveman-shrink MCP middleware is a separate Node package under src/mcp-servers/.

When to reach for it

Agent replies are too long and you want shorter, denser answers without losing technical content.
You want compressed commit messages and one-line PR review comments.
You want to shrink a recurring memory/context file (like CLAUDE.md) so every future session loads a smaller prompt.
You use several agents and want one installer to apply the same style everywhere.

When not to reach for it

You care about total per-session cost: the README is explicit that only output tokens shrink, the skill adds roughly 1–1.5k input tokens per turn, and on already-terse workloads savings can go net-negative.
You want to reduce the actual amount of code the agent writes rather than the words it uses — that is a different tool (see Alternatives).
You need verbose, explanatory output (teaching, detailed rationale), which this deliberately removes.
Your host is one of the few agents the installer does not support.

Maturity signal

The project is recent (created April 2026) and popular, with a very high star count, pushes as of early July 2026, and an MIT license; the installer package sits at an early 0.1.0. Committed benchmarks and evals, a documented "honest numbers" caveat, and a maintainer guide (CLAUDE.md) point to an actively maintained project that is candid about the limits of its headline metric rather than one coasting on a meme premise.

Alternatives

ponytail — use it when the problem is the agent over-building rather than over-talking; it trims the code written, and the two are described as complementary.
caveman-code — use it when you want a whole terse coding agent end to end, not just a skill that compresses replies inside your existing agent.
The default verbose agent — use it when you specifically need long, explanatory answers and are not optimizing for output length.

Notes

Savings apply to output tokens only; input and reasoning tokens are untouched, so whole-session numbers are smaller than the 65% headline and the docs point to a dedicated "honest numbers" writeup.
The README cites a March 2026 paper (arXiv 2604.00025) reporting that constraining large models to brief answers improved accuracy on some benchmarks, as evidence that brevity can help correctness, not just cost.
The wenyan level is an intentional exception to the "keep your language" rule because classical Chinese packs the most meaning per token.

Wiki: JuliusBrussee/caveman