Skip to content

caveman (Julius Brussee)

This is a tool reference

What a tool or project is, who makes it, and what it is for, whether someone else wrote it or I did. Everything I have learned about it lives elsewhere and links back here.

Caveman1 is a skill that tells an agent to answer in fewer words: drop filler and hedging, drop the narration around tool calls, and answer in the shape [thing] [action] [reason]. [next step]. It is one SKILL.md and nothing else, and it is the free layer of a commercial stack2.
Author Julius Brussee
Licence MIT, for the skill
Distribution npx -y github:JuliusBrussee/caveman, or the vendored install.sh
Source github.com/JuliusBrussee/caveman
Site caveman.so
Version read commit ae26f3a, 2026-09-21

What is actually being offered

Five layers share the name, and only the first is prose. The skill is MIT-licensed and local. The proxy and middleware below it compress context on its way to a provider and keep the original for recovery. The platform and enterprise layers above are a gateway with caching, routing and a signed savings ledger, both still in development.

Everything except the skill is about spending less money. The skill is the only layer that changes what a reader sees, and it is installable on its own, with no account.

The part worth having

The skill defines graded levels. lite cuts filler, hedging and pleasantries whilst keeping articles and full sentences. full additionally drops articles and allows fragments, which is the register the name promises. ultra strips conjunctions on top of that. Three further levels answer in classical Chinese.

Two of its rules are about reading rather than counting, and they are where the value is. It mixes in ASD-STE100 Simplified Technical English, the aerospace documentation standard: one idea per sentence, twenty words, active voice, one term per concept, imperatives for instructions. And it forbids narrating tool calls, which is where an agent's reply length mostly comes from.

It also argues against its own folklore, which is the most interesting thing in the file. Invented abbreviations such as cfg and impl tokenise the same as the full word, so they save nothing and cost the reader a decode. Causal arrows are their own token. Mangled verb agreement is the same length as correct agreement. The instruction that follows is to use plain phrasing wherever caveman phrasing is not shorter, which makes full and ultra a smaller change than they appear.

Compression is suspended for security warnings, confirmations of irreversible actions, and anywhere terseness would make an order of operations ambiguous. It applies to replies only: code, comments, commits, documentation and anything written for another human stay in ordinary prose.

Installing it

The installer recognises around a dozen harnesses and writes each one its native form, with opencode amongst them; --with-init additionally drops an always-on rule body into AGENTS.md and the per-agent rule directories. For Claude Code it also installs hook scripts that track the active level and render it into the statusline, which is the only place the level survives being pushed out of context.

Installing the skill by hand is a copy of one file, because a skill is discovered by its SKILL.md rather than declared, the same property the skills installer is built on.

What I took of it, and what I left, is Adopting caveman at lite, without its init block.

The following pages link to this page:

  1. JuliusBrussee/caveman: ultra-compressed communication mode for coding agents, last modified 2026-09-21 ↩

  2. Caveman - the token-efficient stack for agent-native development, last modified 2026-09-21 ↩