How to use the Caveman skill: install it and cut tokens
Caveman is an open-source skill that shortens the answers of AI coding agents to save tokens. It leaves code, commands and error messages untouched and removes only greetings and filler. This guide covers installation, the commands that turn it on and off, the measurements the repository publishes and the tasks it suits.
Translated from Korean. View original

In this guide
- What is the Caveman skill?
- What does it cut, and what does it keep?
- How do you install Caveman?
- How do you turn it on and off?
- How many tokens does it actually save?
- Can it shrink memory files such as CLAUDE.md?
- Which other commands are included?
- Which tasks does it suit?
- What should you check before installing?
- Frequently asked questions
- Sources
What is the Caveman skill?
Caveman is a skill published on GitHub by Julius Brussee. It teaches an agent a set of speaking rules: answer first, put one idea in each sentence and never drop meaning. The name comes from answering briefly, like a caveman.
The repository is released under the Apache-2.0 license and had passed 100,000 GitHub stars as of November 2026. It works in Claude Code and in other agents such as Codex, Gemini CLI and OpenCode.
The example below comes from the repository README. The same answer takes 63 tokens in a normal reply and 20 tokens in a Caveman reply.
- The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle.
- Two sentences of explanation are followed by one sentence of advice.
- New object ref each render, so React re-renders. Wrap the prop in useMemo.
- Only the cause and the fix remain.
What does it cut, and what does it keep?
Caveman does not work by breaking grammar. It fixes the order and length of an answer and leaves the parts that carry information alone. The README says the rules are built on ASD-STE100, the controlled English used in aircraft maintenance manuals.
It does not change the requests you write. The README cites research showing that shortening the user's prompt makes answers worse.
| Rule | What it means |
|---|---|
| Answer first | No greeting, no “let me take a look”, no recap |
| One idea per sentence | Sentences of 20 words at most, in the active voice |
| Meaning stays | Words such as not, never and only stay, along with numbers and units |
| Code stays exact | Code, commands, paths and error messages are not changed by a single character |
| Knows when to stop | Security warnings and irreversible actions are explained in full sentences |
How do you install Caveman?
If you only want shorter answers, install the skill alone. Run “npx skills add JuliusBrussee/caveman -g” in a terminal. To remove it, run “npx skills remove JuliusBrussee/caveman -g”.
In Claude Code you can also install it as a plugin. Type “/plugin marketplace add JuliusBrussee/caveman” and then “/plugin install caveman@caveman”. The plugin starts automatically in every interactive session.
“npx @caveman-ai/cli@latest” is the full install, which adds a local proxy alongside the skill. It brings features such as tool-output compression and needs Node.js 22.13 or later. If you are new to it, try the skill alone first and decide afterwards.
| Install method | Command | What you get |
|---|---|---|
| Skill only | npx skills add JuliusBrussee/caveman -g | The speaking rules for short answers |
| Claude Code plugin | /plugin install caveman@caveman | The skill and automatic start in each session |
| Full install | npx @caveman-ai/cli@latest | The skill, a local proxy and extra features |
How do you turn it on and off?
If it does not start on its own, type “/caveman”. In Codex the command is “$caveman”. To return to the normal voice, say “stop caveman” or type “/caveman off”. “/caveman status” shows the current mode.
There are three levels: the default “/caveman”, the shorter “/ultracave” and “/megacave”, which answers in classical Chinese. “/megacave” is hard to read, so choose between the first two for everyday work.
| Command | The same answer | Tokens |
|---|---|---|
| /caveman | New object ref each render, so React re-renders. Wrap the prop in useMemo. | 20 |
| /ultracave | Inline object prop, new ref, re-render. useMemo. | 14 |
| /megacave | One line of classical Chinese | 13 |
How many tokens does it actually save?
The repository publishes an example run on ten developer questions. Output was 6,983 tokens with no instruction, 4,334 tokens with only “Answer concisely”, 4,119 tokens with “/caveman” and 2,693 tokens with “/ultracave”.
The README treats “Answer concisely” as the real baseline. Against it, “/caveman” cuts 3% more at the median and “/ultracave” cuts 35% more. This is a ten-question example, so your own work may give different results.
The README also lists cases with no saving. The rules themselves cost about 1,160 input tokens, and in tools that bill per request, such as GitHub Copilot, shorter answers do not lower cost. The JetBrains A/B test cited in the README found no measurable quality difference across 86 real coding tasks and 8.5% fewer output tokens.
Total output tokens across ten developer questions.
Can it shrink memory files such as CLAUDE.md?
The “/caveman-compress” command rewrites memory files such as CLAUDE.md in a shorter form. In the README's measurement, five files went from 6,198 tokens to 4,138 tokens in total, and each file shrank by 22.8% to 49.1%. Headings, code blocks and paths stay intact.
Memory files are read at the start of every session, so compressing one once keeps paying off in later sessions. The original is kept as a backup, so read the result and check that nothing is missing. Note that this command sends the file being compressed to your configured model provider.
Which other commands are included?
Besides the commands that change the voice, Caveman includes task commands that write briefly. They suit text with a fixed format, such as commit messages and code reviews.
| Command | What it does |
|---|---|
| /caveman-commit | Writes a one-line Conventional Commit message |
| /caveman-review | Lists one finding per line |
| /caveman-compress | Shrinks memory files and backs up the original |
| /caveman-stats | Shows token usage for the current Claude Code session |
| /caveman-help | Shows every mode and command on one screen |
Which tasks does it suit?
It suits work where you only need to check the result quickly. For work that needs a full explanation, it is better to turn it off. You can turn it off with “stop caveman” and back on with “/caveman” in the middle of a task, so switch as the situation changes.
- Debugging where you only need the cause and the place to fix
- Commit messages and code reviews
- Checking progress in a long session
- Detailed explanations of something you are learning for the first time
- Documents and notices written for other people
- Tools that bill per request
What should you check before installing?
According to the README, the skill on its own runs only on your machine and sends nothing. The one exception is “/caveman-compress”, described above.
The CLI that comes with the full install sends usage statistics by default, such as the commands run and token counts. The README states that prompts, code and file paths are not sent. You can turn the statistics off with “caveman telemetry off”.
Skillshop also has a page introducing Caveman. The guides below cover other ways to reduce tokens.
Frequently asked questions
Is the Caveman skill free?
The whole repository is released under the Apache-2.0 license, so it is free to install. Running “npx skills add JuliusBrussee/caveman -g” in a terminal installs the skill alone.
Does Caveman make answers less accurate?
It shortens wording and leaves code, commands, numbers and error messages as they are. The JetBrains A/B test cited by the repository found no measurable quality difference across 86 coding tasks. Compare the results on your own work as well.
How do you turn Caveman off?
Say “stop caveman” or type “/caveman off” to return to the normal voice. Type “/caveman” to turn it back on.