Tools
Claude Code (or 30+ agents), caveman skill
Price
Free (Apache 2.0 licence)
Time
5 min
Level
Terminal, beginner

You ask your agent why a script crashes, and you get "Great question!", three paragraphs of explanation, a summary and an offer to go further. The useful answer fit in one line. caveman teaches the agent to keep it short: less for you to read, less text for the machine to generate.

What the skill does

caveman is published by Julius Brussee on GitHub, under the Apache 2.0 licence, and is very active (version 4.0.1 released on 13/10/2026). The skill enforces a way of answering, set out in its instruction file:

  • the answer first, no greeting, no "let me check", no recap;
  • one idea per sentence, 20 words at most;
  • words that change the meaning (not, never, only) and numbers always stay;
  • code, commands, file paths and error messages are copied exactly;
  • security warnings, irreversible actions and moments when you seem lost switch back to full sentences.

The README gives an example: a 63-token explanation (tokens are the word pieces an AI model counts) brought down to 20, with the same fix.

What it really saves

This is the interesting part: the repository has an "Honest Numbers" page that qualifies its own claims. What it says, as of 17/10/2026:

  • The skill only shrinks output. It does not compress your messages, the context, the files read or the model's thinking. Input reduction: 0%.
  • Against a plain "Answer concisely.", the median saving measured on 10 questions is 3% for /caveman and 35% for the harsher /ultracave mode. A small sample, measured on an older version of the rules.
  • The rules have a cost: about 1,160 extra input tokens. On short Q&A, one user measured a net loss.
  • Per-request billing: if your tool counts requests rather than tokens (GitHub Copilot, for example), a shorter answer saves nothing.
  • An outside test cited by the repository (300 runs per arm) found 44% fewer output tokens, but a total cost up 0.06%: an agent's bill comes mostly from what it reads, not what it writes.

Since version 4, caveman is also a broader tool (compressing what the agent reads, picking the model automatically) with its own numbers: that is a different product, with an account and telemetry. Here I only cover the skill.

Install the skill on its own

  1. Add the skillWith the skills.sh tool (Node.js required). The README gives this command for the skill alone, without the rest of the tool:
    Terminal
    npx skills add JuliusBrussee/caveman -g
    The tool shows you the list of skills in the repository: tick only caveman if you just want short answers.
  2. Turn it onIn Claude Code, type /caveman if it does not start on its own. To go back to normal, say "stop caveman" or "normal mode".
  3. Remove it if needed
    Terminal
    npx skills remove JuliusBrussee/caveman -g

Example: a maintenance session

You spend the afternoon fixing a client's website with Claude Code. You read every answer: that is where caveman helps most, in reading time more than in money. Start the session like this:

Claude Code
/caveman
The site's pages have been showing without styles since the last deploy. Find the cause and suggest the fix.

For the email to the client explaining the outage, say "stop caveman": a text meant for a human deserves full sentences.

Useful settings

  • Language: the skill's rule is to compress the style, not the language. It answers in your language, and an explicit instruction ("Answer in French") always wins.
  • Measure it yourself: the repository recommends it. Run the same task with and without caveman, and compare usage on your provider's usage page. If it costs more, turn it off for that kind of work.

Limits

  • With a Claude subscription, the main benefit is reading comfort. The effect on your usage limits is not measured by the repository.
  • Very short answers can hide an assumption. If an answer feels too terse, ask "explain in full sentences".
  • Do not use it to write client texts, offers or content: that is not its job.

The skill alone, or the whole tool? The README puts forward a full install (npx @caveman-ai/cli@latest or a curl … | bash script) that changes the configuration of all your agents. That tool sends usage statistics by default (commands, token counts, an install ID, IP address), and its "Auto" mode sends your latest requests to an online service once you sign in. If all you want is short answers, the skill alone is enough: according to the README, it runs on your machine and sends nothing. Either way, read the SKILL.md before installing.

Sources

Cutting an AI agent's cost is mostly about what it reads and which model it uses. That is a setting I tune on the AI agents I build.