Caveman AI: Token Compression for AI Coding Agents
The Problem: AI Agents Talk Too Much
Do you know the feeling? You ask Claude Code a simple question and the answer reads like a school essay. Polite phrases, introductions, summaries, another summary of the summary. You scroll and scroll, even though the actual info could have fit into two sentences.
It’s not just annoying. It also costs money.
With token-based billing models, you pay for every single word the agent spits out. If 70% of it is filler material, it quickly becomes expensive. Julius Brussee looked into this and built a pretty unconventional solution: Caveman AI. An open-source skill that forces AI agents to respond in a compressed manner, with up to 75% less output tokens according to project claims.
What is Caveman AI?
Caveman AI prompts AI coding agents to respond in a radically compressed style. The idea: remove filler words, remove polite phrases, remove hedging, keep technical accuracy. The name says it all, the agent communicates like a caveman.
Instead of:
“Of course! I’m happy to help you fix this error. Let me first explain what’s happening here. The error occurs because the variable was not initialized. Here’s the solution I would suggest…”
You get:
“Bug: Variable not initialized. Fix: Add let count = 0 at line 12.”
Over 27,000 GitHub stars for a plugin that has only existed for a short time. The community celebrates the approach.
How does it work?
Technically, Caveman AI hooks into the workflow of your coding agent as a skill. With Claude Code, the system works with three hooks:
- SessionStart activates caveman mode and writes a flag file
- UserPromptSubmit detects /caveman commands and updates the mode
- A status bar script shows you which mode is currently active
The central control file is skills/caveman/SKILL.md, the single source of truth for the entire behavior. CI workflows automatically synchronize it into all agent-specific directories. So you don’t have to dig into the guts of various configurations. One file rules them all.
The Three Intensity Levels
Caveman doesn’t come as one-size-fits-all. You have three levels:
- Lite: Professionally concise. Full sentences remain, but all superfluous content is removed. Good if you also show the agent to colleagues and don’t want to appear like a Neanderthal.
- Full: The classic caveman mode and the standard. Short fragments, no polite phrases, straight to the point.
- Ultra: Maximum compression. Abbreviations, arrows instead of conjunctions, almost telegram style. Not for everyone, but if you know what you’re doing, extremely efficient.
There are even 文言文 modes (wenyan) that compress in classical Chinese. Whether you’ll ever need that is another question, but it shows how far the concept has been thought out.
In addition to the levels, Caveman offers specialized commands:
caveman-commit
for concise commit messages,
caveman-review
for one-line code reviews,
caveman-compress
for compressing your own input tokens.
Token Savings and Speed
Now it gets exciting, but also a bit tricky. The project claims an output token reduction of about 65 to 75% and a speed increase by a factor of 3. The tool
caveman-compress
is also supposed to save around 46% of input tokens.
But, and this must be said honestly: These numbers come from the project itself. Internal project benchmarks, no independently verified studies. The README even speaks of “100% technical accuracy,” which is a pretty bold claim.
The project landing page cites a paper, “Brevity Constraints Reverse Performance Hierarchies in Language Models,” which is supposed to prove that enforced brevity even improves accuracy. Sounds good, but has not yet been independently confirmed.
Does that mean the numbers are nonsense? Not necessarily. The basic observation that LLMs are unnecessarily verbose and compression works is plausible. But take the specific percentage values with a healthy dose of skepticism. Especially with Agentic AI Workflows, where agents autonomously perform multiple steps in succession, verbose response behavior can become a real cost factor, even moderate savings add up quickly.
Supported Agents and Installation
Caveman supports a whole range of agents: Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, GitHub Copilot, and according to project claims, over 40 more via
npx skills add
in Claude Code, Gemini CLI, and the local Codex setup, auto-activation works natively. With other agents, you have to manually add an always-on rule, not a big effort, but good to know.
The installation itself? An npx command, wait a bit, done.
Auto-Clarity: When It Gets Serious
One of the more thoughtful aspects: Auto-Clarity. The agent automatically switches back to the normal, detailed language style when it comes to security warnings, irreversible actions, or multi-step sequences.
With an rm -rf, you definitely don’t want a caveman answer, but a clear, unambiguous warning. Code blocks, Git commits, and PR descriptions also remain unchanged, technical terms are kept exactly.
Especially when working with more complex setups, such as Multi-Agent Systems, where multiple agents work together, this security layer is important. An agent that responds cryptically in critical moments would be a real problem.
Is Caveman AI Worth It?
Caveman AI solves a real problem. LLMs are verbose, it costs time and money, and compressed answers are often actually more readable. The concept is elegant, the implementation thoughtful, and the GitHub community sees it similarly.
Nevertheless: The performance claims have not yet been independently confirmed. The Ultra mode can become too concise with complex explanations, causing you to lose context. And if you use the agent as a learning tool, you might not want compressed answers at all.
For experienced developers who know exactly what they need and use their coding agent as a tool rather than a tutor, Caveman AI is definitely worth a look. Start with the Lite mode and work your way up. In the worst case, you turn it off again.
Our Sources
- caveman/README.md at main · JuliusBrussee/caveman
- GitHub - JuliusBrussee/caveman: 🪨 why use many token when few token do trick - Claude Code skill that cuts 65% of tokens by talking like caveman
- caveman/CLAUDE.md at main · JuliusBrussee/caveman
- caveman/caveman.skill at main · JuliusBrussee/caveman
- So Expensive, A Caveman Can Do It | Hackaday
- caveman/CLAUDE.original.md at main · JuliusBrussee/caveman
- caveman/hooks/README.md at main · JuliusBrussee/caveman
- caveman/skills/caveman/SKILL.md at main · JuliusBrussee/caveman
