Hey, took a look at that caveman skill everyone's sharing (https://github.com/JuliusBrussee/caveman), and I wanted to share a couple of thoughts. So the pitch is talk like caveman, cut 65% of tokens. and the 65% is real. I checked their benchmark code, they're calling the actual Anthropic API, running 10 prompts × 3 trials, measuring output tokens. average comes out to 65%, range 22–87%. But here's the thing: it only compresses output token. and in an agentic session, output is a tiny part of what you actually pay for. I ran the numbers on my own history, 186 sessions, 14k assistant messages: • input: ~1.97B tokens • output: ~3.3M tokens • that's a 591 : 1 ratio at Antrophic API pricing that's roughly $3,600 input vs $250 output. so output is ~6% of the bill. cutting 65% of that saves you like 4% total. To their credit they do say it in the README: "biggest win is readability and speed, cost savings are a bonus." . I believe the interesting part is actually

caveman-compress

, which rewrites your AGENT.md into caveman-speak so your model read fewer tokens every session. that one hits the input side, which is where the real money is. Haven't tested it yet but it seems interesting.

@modern-fireman-99755 Interesting findings! I guess we can't compress the input with a smaller llm? I wonder what caveman code would look like. It could almost be interesting to run a full benchmark and compare the price / time / performance tradeoffs. But if the gain is epsilon, maybe not... Or maybe just for a blog post? @limited-student-10747

> I guess we can't compress the input with a smaller llm? Actually, yes. There are people at EPFL researching these things. Specifically, compressing the prompt with a small, cheap model before passing it to a bigger one. They claim that they are reducing cost (input token) while maintaining the same accuracy. They’ve published some papers on it, for example: https://arxiv.org/html/2511.12281v2