I think that everyone who uses AI recognizes the following pattern; you ask an LLM a simple question and it answers like it's writing a blog post: introduction, context, three examples, a closing summary. Fine for a first read, expensive when you're chaining calls or running an agent loop all day.
The trick to avoid this is called "caveman prompting". You tell the model to drop articles, pleasantries and filler, and answer in short, blunt fragments. It sounds silly. But it works up to a point.
A first attempt: just say "be concise"
Most people's first instinct is a one-line system prompt:
Be concise. No fluff.This already gets you a good chunk of the savings. In benchmarks I've seen floating around, a plain "be concise, return structured output" instruction accounts can already give you a nice reduction. It's the cheapest fix and most people stop here, which is reasonable.
The caveman approach
The caveman skill takes this idea further and transform it into a token-saving solution. Just drop it as a skill/plugin into Claude Code, Copilot, Cursor, Codex, Gemini or one of a few dozen other agents. One install command:
npx skills add JuliusBrussee/cavemanThe skill file tells the agent to drop articles, filler and pleasantries, answer in fragments, and use short synonyms ("big" instead of "extensive", "fix" instead of "implement a solution for"). Code, commands, error strings and symbols are explicitly left untouched, byte-for-byte.
It isn't all-or-nothing. There are several modes you switch between with a slash command:
/caveman litedrop filler and hedging, keep full sentences. Professional but tight./caveman(default, "full") drop articles too, fragments are fine, short synonyms./caveman ultrabare fragments, standard abbreviations (DB, auth, fn), arrows for causality.
Remark: The caveman skill only shrinks what an agent says. If you want that it shrinks everything, have a look at the caveman-code skill.
What I saw when I actually tried it
I installed it on a few real sessions rather than trusting the README numbers. The reduction was real, but nowhere near the headline 65%. Depending on the task, I landed somewhere between 15% and 30% fewer tokens.
The variance made sense once I looked at what I was doing in each session. Short, back-and-forth debugging sessions barely moved the needle, the ~1-1.5k token overhead of the skill instruction itself was eating a good chunk of whatever it saved. Longer sessions with more explanatory answers, code reviews, architecture discussions, got closer to the 30% end.
Remark: I'd treat the 65% figure as a ceiling reached under fairly specific conditions, not a number to plan a budget around. If you're chaining a lot of short calls, measure it yourself before assuming the savings are there.
The author has built a small family of tools around the same idea: a full terminal coding agent that's caveman by default, a memory layer that stores session context compressed in the same grammar so it stays smaller across sessions, and even a fine-tuned model where the compression is baked into the weights instead of a system prompt. Worth a look if you're building agent-heavy workflows and the per-turn overhead of a text instruction actually matters to you.
More information
- JuliusBrussee/caveman: why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
- I Benchmarked the Viral "Caveman" Prompt to Save LLM Tokens
- What is Caveman Prompt? Reduce LLM token usage
- Token Saving, and Caveman - DEV Community
- JuliusBrussee/caveman-code