I stumbled across a GitHub project called Caveman, and the idea made me laugh. It’s an AI skill with a simple goal: make the assistant talk less.

Not by making it dumb, or having it say “me fix code” all day. It strips out the usual AI pleasantries and filler while keeping the important technical details intact.

Instead of:

Sure! I’d be happy to help you investigate that issue. Based on the code you’ve shared, it looks like…

You get something closer to:

Bug in auth middleware. Check token expiry comparison.

That’s pretty much the whole idea, and I kind of love it.

Less talking, fewer tokens

The interesting part is that this can actually save output tokens. AI responses are made of tokens, and many APIs charge for the tokens a model generates. Shorter answers can mean lower output costs, especially if you’re using an assistant constantly while coding.

Caveman’s rules are more thoughtful than the name suggests. Keep code, commands, paths, numbers, errors, and negations exact. Remove ceremony, not meaning. It even says not to force fake caveman grammar.

Of course, shorter isn’t always better. Ambiguous instructions or missing technical details can cost much more than the few tokens you saved. And the skill itself takes some context to load, so there’s no guarantee every single interaction uses fewer total tokens.

Still, there’s something funny about needing an entire set of instructions to tell an AI to stop talking so much.

Less fluff. Same useful answer. Potentially fewer tokens.

Check it out on GitHub: JuliusBrussee/caveman.