Caveman: FREE Tool Cuts Tokens by 69%!

**Julian Goldie** (0:00)
So today, we're going to be looking at a new way to cut 65% of tokens with Claude Code. So if you're using Fable 5, this is an easy way to cut the amount of tokens you use, but still get the most out of it. So you could use it even more. I mean, imagine if you can code 65% more with Fable 5, how much more would you get done? 65% more. So that's why we are using this system to reduce the amount of tokens that we use when we are running AI agents. And you can also use this with any other agent. You know, it's not just for Claude Code. You can also run it with Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and it's free and ready to go. So you can see right here. You also might say, how does this work? Well, I'll run through that in a second. And if you want to indicate how to install it, you can just copy a basis command and give the whole setup to Claude Code so that it can run with Fable 5
And the great thing about this is you can install it for one agent or even for 30 different agents, which is pretty insane. And the way this works is basically you talk to your agent like a caveman.
And when we actually tested on Fable 5, it used 69% fewer output tokens. I'll show you the example test in a minute. It was 37% cheaper and every answer was still pretty good. So here's a problem with, for example, something like Fable 5, it uses up a lot of tokens. When it stops being free on the subscription then you have to pay for API. It's gonna get quite expensive if you're not careful. As a caveman is a free little skill that fixes the talking, not the thinking. And you install it once and then your agent basically answers like a smart caveman, short, blunt and right. And then the way this works is basically, the idea is that when AI answers you, you are using up tokens for every word it says. And those are output tokens. And on the smartest models, they're quite expensive. They use up a lot of tokens. With caveman, it's a set of instructions your agent reads once at the start of every chat. And the instructions say, for example, drop the filler words. No, like for example, sure, I'd be happy to help. No, basically and actually no. Three paragraph warmups. You say the thing, you give the fix and it stops talking. And there's basically three rules inside caveman to make it use less tokens. So number one is code is sacred, right? The code blocks commands, file paths, error messages are never touched. Then also it knows when to speak up. So obviously, sometimes it will need to speak more in situations related to security. And so this way it knows, okay, during those situations, we're going to speak more. And also it can speak multiple different languages. Well, if you, for example, write in Portuguese, it will reply like a caveman in Portuguese. It compresses the style, but not the actual meaning. And so the way this works is like it thinks exactly the same. It's still thinking like it would normally, but the difference is the filter afterwards. So all the filler words that come back to you are reduced. And so you use less output tokens when it's replying back to you, when it's coming back to you inside the chat.
Also, you might say, okay, you know, aren't shorter answers when it's replying to you, like making it worse? Well, basically, apparently sometimes this can be the opposite. So March 2026 research paper tested 31 models and found forcing brief answers improved accuracy by up to 26 points on some benchmarks. And in my own five prompt test, every caveman answer contained the same fixes along the version, right? So it was still working just as well. It's just like it replies to less words and it's a bit shorter and more straight to the point. And if you're wondering, okay, what it actually is and what happens in every message?
Basically, it's just a text file full of rules. So that's the whole secret. It's not like an app or a new model or a middleman or anything like that. It's just like a written instruction, a skill that tells your agent how to speak, nothing more. Now the installer basically copies that file into your agent. So you rent the one line install, it looks in machine, finds every agent you have. And then with Claude Code, Codex, Gemini, 30 others, it drops the same rules file into each one skill or plugin folder in the format that agent expects.