01What Is Caveman Tool and How It Cuts Tokens

Caveman tool is a free, open-source skill that cuts the number of output tokens your AI agent uses. It does this by adding a set of instructions that make the agent reply like a smart caveman: short, blunt, and to the point.

The tool is not an app or a new model. It is just a text file full of rules that your agent reads at the start of every chat. The rules tell the agent to drop filler words, pleasantries, and long warmups.

Instead, it gives you the answer directly.

This reduces the output tokens because every word the AI writes costs tokens. With fewer words, you use fewer tokens. In tests with Fable 5, it saved 69% of output tokens and 37% of total costs.

02Key Features of the Caveman Tool

Caveman tool has several features that make it useful for anyone running AI agents.

  • One-command install: You copy and paste a single command to set up the skill for one agent or up to 30 different agents.
  • Works with many agents: It supports Claude Code, Codex, Gemini, Cursor, Windsurf, Klein, Copiler, and more.
  • Three reply modes: You can choose between light, full, and ultra to control how short the replies are.
  • Protects crucial content: Code blocks, commands, file paths, and error messages are never shortened or altered.
  • Multi-language support: If you write in Portuguese, it replies like a caveman in Portuguese, compressing the style but not the meaning.

You can also turn it off anytime with the /caveman command. Bonus commands like caveman commit, caveman review, and caveman compress can reduce tokens even more.

03How to Set Up Caveman Tool Step by Step

Setting up caveman tool is simple and takes just a few minutes. Here is how you do it.

  1. Get the one-line install command from the project’s GitHub page.
  2. Copy the command into your terminal.
  3. Run the command. It will scan your machine and find all your agents.
  4. The installer drops the rules file into each agent’s skill or plugin folder in the format that agent expects.
  5. Start a new chat with your agent. The rules are now active.

If you are not technical, you can give the GitHub link to Claude Code and let it set up the tool for you. That way, you do not need to touch the terminal yourself.

You can also install it inside your agent operating system if you use multiple agents like Hermes, OpenClaude, Claude, or Antigravity. That way, the token savings apply across all of them.

04How to Use Caveman Tool in Daily Work

Using caveman tool is as simple as using your normal AI agent, but you get shorter answers. For example, if you ask why your React component is re-rendering, the agent will read the caveman rules and reply with the fix directly, without a long intro.

The rules tell the agent to follow a pattern: thing, action, reason, next step. So you get a clear answer that is easy to skim.

You can change the reply style by switching between light, full, and ultra. Light gives a slightly shorter answer, full is shorter, and ultra is the shortest. To change the setting, you just modify the mode as shown in the episode.

You can also use bonus commands like caveman commit to compress commit messages, caveman review to get quick code reviews, and caveman compress to shrink long files. These commands help you save even more tokens.

05Test Results: 69% Token Savings and More

In this episode, Julian Goldie ran five real questions from GitHub through Fable 5 using the caveman skill. Every single test saved a significant number of tokens.

For example, a React re-render bug question normally used 1,349 output tokens. With caveman, it used only 324 tokens. That is a huge reduction.

The average savings across the five tests was 69% on output tokens. This translated to about 37% lower total cost because output tokens are only part of the cost.

Julian also checked the answers. All of them were correct. The shorter replies did not reduce accuracy.

In fact, a March 2026 research paper tested 31 models and found that forcing brief answers improved accuracy by up to 26 points on some benchmarks.

06Tips, Limits, and Mistakes to Avoid

While caveman tool is powerful, there are some important tips and limits to keep in mind.

  • Do not use it for security-sensitive tasks: The rules include a clause that allows the agent to speak more when the situation involves security. Trust that.
  • It works best for coding tasks: The tool is designed for AI agents like Claude Code, not for general chat bots.
  • You can turn it off: If you need a longer explanation, use /caveman to disable it temporarily.
  • Choose the right mode: Ultra might be too short for complex problems. Start with light or full.
  • Install once, save repeatedly: The rules file is read at the start of every chat, so the savings compound across all your sessions.

One mistake is thinking that shorter answers mean worse answers. But tests show otherwise. The agent still thinks the same way; it just writes fewer words.

07Who Should Use Caveman Tool and Why It Matters

Caveman tool is for anyone who uses AI agents for coding, especially those on subscription plans with limited tokens or those paying per API call.

If you use Fable 5, you will soon have to pay for API usage after the subscription drops on July 7. Reducing tokens with caveman saves you a lot of money.

Even if you use other agents like Claude Code, Codex, or Gemini, you can benefit from the same savings. The tool is free, open source, and easy to install.

By cutting tokens by 69%, you can do 65% more work with the same budget. That is a big deal for developers and teams who rely on AI agents daily.