caveman: The Claude Code Skill That Cuts 65% of Your Output Tokens
TL;DR
- caveman strips articles, pleasantries, and hedging from an AI agent's replies while keeping code blocks, error strings, and technical terms exact.
- Measured output reduction averages 65% across a 10-prompt benchmark, ranging from 22% to 87% depending on the task.
- It works past Claude Code alone: Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and more than 40 other agents via npx skills.
- No telemetry, no accounts, no backend. Once installed, it makes zero network calls; the compression happens entirely as a prompt plus local hooks.
Every reply an AI coding agent sends back costs tokens, the same way every prompt sent to it does. Most people notice the input side, since that’s the part they typed. Fewer notice that a verbose, hedge-heavy answer costs just as much on the way out, and that cost compounds every time the model has to loop back through its own growing reply to generate the next word.
That compounding is what a 19-year-old developer named Julius Brussee set out to cut with a skill called caveman.
Why does a longer AI answer cost more than it looks like?
Large language models process text as tokens, not raw characters, and generation is autoregressive: the model produces one token, appends it to what’s already been said, and reruns its full forward pass on the extended sequence to pick the next one. A short reply does that loop a few dozen times. A long, hedge-padded reply does it hundreds of times, and each pass costs compute regardless of whether that particular sentence needed to exist.
Streaming output hides this from most users, since text appears to flow in real time as if the model were narrating its thoughts live. What’s actually happening is a fast repeat of that same generate-and-rerun cycle, and running a high-reasoning model like Opus or GPT-5.5 in extended thinking mode multiplies the token count further, which shows up directly on the bill for anyone paying per API call.
What does caveman actually do?
caveman rewrites an agent’s response style: articles, greetings, and hedging language get cut, while code blocks, error messages, variable names, and every technical detail stay exactly as they were. Measured against default verbose replies across a 10-prompt benchmark, it cuts an average of 65% of output tokens, with individual results ranging from 22% up to 87% depending on how padded the baseline response would have been.
It also rewrites memory files like CLAUDE.md once, at setup, cutting roughly 46% of the input tokens spent re-reading that file every session afterward. That’s a one-time edit that keeps paying off on every future prompt.
Who built it, and why does that matter?
Brussee is a Data Science and AI student at Leiden University and a founding engineer at Stacklink, an enterprise RAG platform. Caveman isn’t his only project: he’s also building Revu, a local-first macOS study app built around FSRS spaced repetition. The pattern connecting both projects is the same instinct, an aversion to unnecessary token and time expenditure by human or algorithm, wherever it shows up.
Caveman went viral fast after launch, picking up roughly 4,100 GitHub stars in its first three days. That kind of curve usually means the problem it solves was already bothering a lot of people who hadn’t found a name for it yet.
How do you install and run it?
Installation is a single command: npx skills@latest add JuliusBrussee/caveman, which works for Claude Code, Codex, Gemini CLI, and agent-rule formats for Cursor, Windsurf, Cline, Copilot, and more than 40 other tools. Once installed, typing /caveman in the agent’s input activates it for that session, with intensity levels ranging from lite to full to ultra, plus classical-Chinese-inspired wenyan variants for maximum compression.
There’s no telemetry, no account creation, and no backend service involved. After install, the skill makes zero network calls: the compression logic lives entirely in the prompt and local hook scripts, and a /caveman-stats command reads token-usage numbers already sitting on disk rather than phoning anything home.
The tradeoff is genuinely small for anyone billed per token or per API call: shorter replies, same technical content, and a noticeably shorter wait for each response to finish generating.
References
You might also like