Caveman skill for Claude

Caveman started, by its author's account, as a joke in April 2026 and now has more than 100,000 stars on GitHub. The idea is simple: a coding agent's replies carry greetings, plans announced in advance and recaps, and all of that costs tokens. Caveman strips them. It is published by Julius Brussee under the Apache 2.0 license; Anthropic does not publish it. For how skills load, see how Claude Skills work.

What the skill does

The README describes caveman as a voice with fixed rules:

• Answer first: the thing, the action, the reason, the next step. No greeting, no "let me", no recap.
• One idea per sentence, at most 20 words, modeled on the controlled English used in aircraft maintenance manuals.
• Meaning stays: articles can go, but "not", "never", "only", numbers and units never do.
• Payload untouched: code, commands, paths and error messages are kept character for character.
• Knows when to stop: security warnings, irreversible actions, step-by-step instructions and confused users get full sentences.
• Your prompts are never rewritten.

There are stronger modes, /ultracave and /megacave, plus helpers such as /caveman-commit for one-line commit messages, /caveman-review for one finding per line, and /caveman-compress for shrinking memory files like CLAUDE.md.

What the numbers say

The README publishes the author's own measurements, including cases where caveman loses. On ten developer questions, plain "Answer concisely." used 4,334 output tokens, /caveman 4,119 and /ultracave 2,693. The README notes that newer models already respond well to "be concise", so the default mode adds only about 3% savings at the median, while the rules add about 1,000 input tokens to every call.

When to use it

• Long Claude Code sessions where chatty replies fill the context.
• Readers who prefer terse answers and can live with dropped articles.

It is the wrong tool for prose anyone else will read. For that, see Humanizer. To cut context use more broadly, read our chapter on context engineering and the Claude Code cheat sheet.

How to install it

npx skills add JuliusBrussee/caveman -g

Type /caveman if it does not start on its own, and say "stop caveman" to go back. As a Claude Code plugin that starts in every session:

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

Example request

/caveman
Why does [component or function] [problem you see]? Show the fix.

Limits

• Savings over a plain "be concise" instruction are small in the default mode, by the author's own count. The bigger cuts come from /ultracave.
• If you pay per request instead of per token, such as GitHub Copilot premium requests, the README says to skip it.
• The repository also offers a separate command-line proxy. The README says the skill sends nothing, but the CLI sends usage stats by default (commands, token counts, an install ID, OS and IP). Turn them off with caveman telemetry off or DO_NOT_TRACK=1.