Talk caveman better
Or how to teach AI to use fewer tokens. LLMs blabber incessantly and that costs pennies, or tokens, the currency of AI.
Caveman
Julius Brussee and Caveman
Who still remembers the days of pay-as-you-go phone calls? It wasn’t that long ago. Those days are probably gone for good, but instead we now have AI tokens. It’s the same idea, but for using an AI subscription. Most people start with the free version, and the more you use the AI, the sooner you’ll sign up for your first subscription. If you’re involved in software development, you’ll quickly upgrade to a Pro account, and even then you’ll soon hit the limit on the number of tokens you’re allowed to use. What could possibly go wrong?
Imagine you’re developing an app; the tokens fly through at an incredible rate: “You’ve reached your limit. Bye" It then quickly becomes expensive, and not just a little bit either. First, you’ll look into how you can use fewer tokens, and that might work a bit, but it won’t make a dent in the budget. But help is on the way, because Julius has come up with a solution: get rid of that chatty language used in those chat apps.
Dutch student Julius Brussee did what AI companies "overlooked": he asked a simple question. "Why use lots of tokens when a few tokens work just as well?"
His answer: Caveman — a Claude Code skill that reduces the verbosity of AI responses by up to 75% by forcing models to communicate like a smart caveman. No filler. No pleasantries. Just the technical core.
Clever. The repository (github.com/JuliusBrussee/caveman) went viral: 63k stars, 3.5k forks. The concept is surprisingly simple: strip articles, hesitations and filler from AI responses whilst retaining every ounce of technical accuracy. A response that previously cost 100 tokens now costs 25.
So? Does it work? Caveman is no joke. It’s a token-efficiency revolution. For AI agents handling thousands of interactions, the savings add up quickly. Companies running fleets of coding agents saw immediate cost reductions. The skill supports multiple intensity levels — lite, full, ultra — and even classic Chinese variants for non-English workflows.
The BNR article (the one on the radio) sees it as a paradox: a teenager beat AI giants with an approach so obvious that nobody had thought of it. Whilst Silicon Valley was chasing bigger models, Brussee asked: what if we just stopped wasting tokens on “I’d be happy to help” and “Let me think about that”? Now you might also think: gosh, those guys make their money from tokens, why would they want less? But anyway, everyone sees it differently.
The viral success of Caveman proves a fundamental truth about AI development: sometimes the biggest breakthroughs aren’t new algorithms, but the removal of what was never needed in the first place. Brussee didn’t build a better model. He built a better way to use the models we already have. And that is, at the same time, one of the remarkable findings: with all those billions from those companies, they are finding it increasingly difficult to innovate. More is happening in the field of AI development in the open-source world than anywhere else.
Anyway. We’ve tried it out and it works brilliantly. Thanks Julius, you’ve already saved us a lot of pointless tokens.
The repository now supports Claude Code, Codex and multiple agent frameworks. The message is clear: in the age of AI abundance, the real scarcity is attention — and Caveman delivers that, one token at a time.