Cut Claude Code output tokens with the caveman skill

Coding agents love to explain. You ask why a component re-renders and get three paragraphs, a greeting, and a recap. Those words cost tokens, and they bury the answer. JuliusBrussee/caveman is a skill that tells the agent to answer short and exact while leaving code and commands untouched.

The README's own example: a 63-token explanation of a React re-render becomes "New object ref each render, so React re-renders. Wrap the prop in useMemo." at 20 tokens.

What it does and what it does not

Be clear about the claim. The voice rule shortens what the agent says. The README is open that agent sessions are mostly code and tool calls, which the voice leaves alone, and that JetBrains measured 8.5% fewer output tokens with no measurable quality loss on 86 coding tasks. It also notes that "Answer concisely." is the real baseline, since newer models already know it. On ten dev questions the README shows 6,983 output tokens with no instruction, 4,334 with "Answer concisely.", 4,119 with /caveman and 2,693 with /ultracave. Those numbers were measured on older rules, so treat them as a guide, not a promise.

Install the skill only

If you only want shorter answers, install the skill for every agent on your machine:

npx skills add JuliusBrussee/caveman -g

Install in Claude Code

Claude Code can add it as a plugin from the shell:

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

Or from inside a Claude Code session, type these where you talk to Claude:

/plugin marketplace add JuliusBrussee/caveman
/plugin install caveman@caveman

Install everything

The newer 4.0 release is bigger than the voice. One command finds your agents and shows what it will change before you press Enter:

npx @caveman-ai/cli@latest

It needs Node.js 22.13 or newer. The README lists extra rocks beyond the voice, such as shrinking logs and test output before the model reads them, and a model picker called Auto. Read what each one changes before you accept it. If you want to back out, the README gives npx -y github:JuliusBrussee/caveman -- --uninstall, and npx skills remove JuliusBrussee/caveman -g for the skill alone.

For Gemini CLI, the README uses gemini extensions install https://github.com/JuliusBrussee/caveman. There is also a one-line install script, but piping a script to a shell deserves a read of the script first.

Use it

Type /caveman ($caveman in Codex) if it does not start on its own. /ultracave is even shorter. Say stop caveman to go back to normal. The README says the voice drops itself for security warnings, irreversible actions and confused users, so those get full sentences.

Check it on your own work

Do not trust anyone's benchmark, including the README's. Run one task twice:

Run <task> once normally and once with caveman on. Compare output tokens and confirm the code changes were identical.

That prompt is the one on our use-case page. If the diffs match and the tokens drop, keep it.

Caveats

• If you pay per request instead of per token, for example with GitHub Copilot, shorter answers save nothing. The README says so.
• The skill adds roughly 1,160 tokens of rules to your input, per the README.
• The README warns that rewriting your own prompt in caveman style makes answers worse, so the skill does not touch your prompts.
• We followed the README and have not run this end to end for this post.

The step-by-step version is on our use-case page: cut your coding agent's token use with one skill. For other ways to trim spend, see the best AI repos for cutting coding agent token usage.

Want the setup commands for 50 AI repos people actually ship with? Get the free AI Repos That Actually Ship pack: https://dedupedai.com/ship-pack/?utm_source=devto&utm_medium=article&utm_campaign=shippack