caveman

Why

"Why use many token when few token do trick."

The joke is doing real work. An agent's default register is padded — hedges, restatements, a preamble before the answer — and you pay for every word of it twice, once in output and again when it re-reads the transcript next turn. This strips the register and keeps the content:

Normal, 69 tokens: "The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object." caveman, 19 tokens: "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."

Nothing was lost. The second one is arguably the better answer.

Version 2 goes after the other half of the bill — what the agent reads — with a proxy that compresses context before the provider call and restores it byte-exactly. The repo reports 33.2% fewer provider-counted input tokens on a pinned benchmark, which is the number worth checking against your own usage rather than taking on faith.

Setup

The skill on its own, which is the part that needs no proxy and no account:

git clone https://github.com/JuliusBrussee/caveman
cp -r caveman/skills/caveman ~/.claude/skills/caveman

Or both halves, including the input-side proxy:

npm install -g @caveman-ai/cli && caveman setup --install
caveman claude

Watch out for

Licensing is split — MIT on the CLI, BSL-1.1 on the proxy runtime. If you're putting this anywhere near work code, read which half you're actually running.

It changes the register everywhere, including where you don't want it. Compressed prose is right for a diagnosis and wrong for a commit message, a PR description, or anything a colleague reads. Turn it off before you generate text that leaves your terminal.