Transcription
Telling Claude Code to talk like a caveman might actually make it smarter, as well as save tokens. So, this is the caveman repo, and what might have started as a meme on Reddit might have some actual legs.
You can see here that by having Claude Code use this repo, which strips out all the filler words when it gives you responses, you're able to save 60, 70, 80% of your output tokens on the text prompt itself. Not total token usage, but the actual prompt it gives you.
But even more importantly, this idea of making your large language model concise might be linked to better outputs. This study that came out last month pretty much said that it showed that when we had larger large language models, think open-weight models of 400 billion plus parameters, when we told them to be more concise, they actually gave better outputs and significantly.
So now the study is linked in this caveman repo if you want to take a look at it. But the caveman repo itself is very, very simple. It's simply a set of skills telling Claude Code to cut all the filler. And it has different levels. So you can be an ultra caveman, a full caveman, or a light caveman.
And it's also important to note that this doesn't change how Claude Code works under the hood. All the code is the same. All the thinking is the same. It's just the actual text outputs.
So, as the great American philosopher Kevin Malone once said, "Why use many word when few do trick?"