GitHub radar
Caveman: Reduce your agent token cost with this one skill!
A skill for Claude Code and 30+ AI coding agents that teaches your AI to avoid unnecessary words like preambles, summaries, and explanations, reducing the number of output tokens by an average of 65%.
Caveman is a skill for Claude Code and 30+ AI coding agents that teaches your AI to avoid unnecessary words like preambles, summaries, and explanations. In our benchmark tests on 10 different tasks, we found that it reduces the number of output tokens by an average of 65%. For example, it reduced 15,704 tokens down to just 121 tokens (a 130x compression ratio). Caveman works either as a transparent proxy or as a Claude Code plugin. We’re upfront about its limitations: it only affects the number of output tokens, and it introduces about 1–1.5k additional input tokens per turn, which means that for shorter conversations it may actually increase the total number of tokens. Its primary advantage is that it makes interactions quicker and easier to read.
Why a vibe-coder should care
Shorter responses can save you a lot of API calls if your AI tends to have lengthy conversations or make frequent use of the web browser. And shorter responses are also easier to scan quickly, which saves time and lets you get on with your work sooner.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the Caveman skill from https://github.com/JuliusBrussee/caveman — pick the right method from the README (npm install or npx skills add), connect it to my Claude Code setup, and confirm it's working. Ask me if you need any permissions.
Runs on a regular laptop — no GPU needed.
Open on GitHub▌ More finds