Does the caveman skill really cut Claude token usage by 65–75%?

Those figures describe compression of prose replies, and they vary by source — the repo says 65%, blog posts say 75%. The repo's own controlled benchmark measured 33.2% fewer input tokens, and that tested the separate proxy product. The skill saves real output tokens, but output is usually the smaller share of a coding session's bill.

Is the caveman skill an official Anthropic skill?

No. It's a community project by Julius Brussee, unaffiliated with Anthropic. Its popularity is real — 102,000+ GitHub stars and a v2.4.0 release in late August 2026 — but nothing about it is official.

Will caveman mode mangle code or error messages?

By its own rules, no. Code blocks, commands, API names, numbers, and error strings stay verbatim, and anything written for other humans — commits, docs, PR text — reverts to normal prose. Compression applies only to Claude's conversational replies.

Is the caveman skill free to use?

The skill itself is MIT-licensed, free for any use. The Caveman Proxy and compression engine in the same repo are BSL-1.1: free to self-host for yourself, but a commercial license is required to offer them as a hosted or embedded service.

Skill review · by Julius Brussee (JuliusBrussee)

Caveman Claude Skill

The caveman skill for Claude makes the agent answer like a terse caveman — articles, filler, pleasantries, and hedging deleted; code, commands, and error strings kept exact — to cut output tokens. It ships from JuliusBrussee/caveman, a repo by Julius Brussee that grew from a one-file joke into a 102,000-star product suite: the MIT-licensed skill plus a BSL-licensed proxy that compresses what the agent reads. The one thing to know before installing: the skill compresses only what Claude says, and in real coding sessions input tokens dominate the bill, so the advertised 65–75% savings describes prose output, not your invoice. The repo's own pinned 54-run benchmark — which measures the proxy, not the skill — found 33.2% fewer input tokens. The skill installs in one command, offers six intensity levels from lite to classical-Chinese wenyan, and carves out security warnings and destructive operations, where it drops the act and writes plainly. It works as described, and it is smaller than its star count suggests.

View the sourceMIT licensed · 102,194★ · facts checked September 1, 2026

By Robert MacKelfresh

Founder, Automation Squad · September 1, 2026

Install

1Claude Code, skill only (MIT part, no hooks): run `npx skills add JuliusBrussee/caveman` in your terminal.
2Or install it as a Claude Code plugin: `claude plugin marketplace add JuliusBrussee/caveman` then `claude plugin install caveman@caveman`.
3Or run the full installer, which also wires Claude Code hooks and a statusline and detects other agents (Node 18+): `curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v2.4.0/install.sh | bash` (PowerShell variant in INSTALL.md).
4In a session, type `/caveman` to activate if it doesn't trigger on "be brief" or "caveman mode"; switch levels with `/caveman lite|full|ultra|wenyan-full`, disable with `/caveman off`.
5Claude.ai (no Claude Code): the skill uses the standard SKILL.md format — download the repo's `skills/caveman/` folder, zip it, and upload it under Settings → Capabilities → Skills.

What it does

Output compression: Rewrites Claude's prose replies in compressed caveman-speak: drops articles, filler (just/really/basically), pleasantries, and hedging; allows fragments; prefers short synonyms. Its own target pattern: "[thing] [action] [reason]. [next step]."
Six intensity levels: lite (no filler, full sentences), full (classic caveman, the default), ultra (one word when one word enough), plus three wenyan modes that answer in classical Chinese. Switch with /caveman lite|full|ultra|wenyan-full; off with /caveman off or "normal mode".
Technical exactness rules: Code blocks, commands, API names, error strings, numbers, and units are never altered, and negations (not/never/no/only) are never dropped. It also bans invented abbreviations (cfg/impl/fn) and arrow glyphs because tokenizers split them anyway — zero tokens saved.
Auto-Clarity failsafe: Automatically suspends caveman-speak for security warnings, irreversible-action confirmations, and multi-step sequences where compressed grammar could be misread, then resumes after the clear part is done.
Boundaries: Anything persisted outside the chat — code comments, commit messages, docs, issue and PR text — is written in normal prose. The mode persists per session until you turn it off.

Common questions

Does the caveman skill really cut Claude token usage by 65–75%?: Those figures describe compression of prose replies, and they vary by source — the repo says 65%, blog posts say 75%. The repo's own controlled benchmark measured 33.2% fewer input tokens, and that tested the separate proxy product. The skill saves real output tokens, but output is usually the smaller share of a coding session's bill.
Is the caveman skill an official Anthropic skill?: No. It's a community project by Julius Brussee, unaffiliated with Anthropic. Its popularity is real — 102,000+ GitHub stars and a v2.4.0 release in late August 2026 — but nothing about it is official.
Will caveman mode mangle code or error messages?: By its own rules, no. Code blocks, commands, API names, numbers, and error strings stay verbatim, and anything written for other humans — commits, docs, PR text — reverts to normal prose. Compression applies only to Claude's conversational replies.
Is the caveman skill free to use?: The skill itself is MIT-licensed, free for any use. The Caveman Proxy and compression engine in the same repo are BSL-1.1: free to self-host for yourself, but a commercial license is required to offer them as a hosted or embedded service.

Before you install

Not an Anthropic product. Community skill by one developer, despite the six-figure star count.
The savings numbers don't agree with each other: the repo tagline says 65%, third-party blog posts say 75%, and the repo's own pinned benchmark measured 33.2% — and that benchmark tests the separate Caveman Proxy, not the skill. No published benchmark isolates the skill's effect on a real bill.
The skill compresses output prose only. In typical Claude Code sessions, input tokens (files read, tool results, context) are most of the cost, so the skill alone moves your spend less than the headlines imply.
The repo is no longer just a skill — it's a mono-repo carrying a proxy, CLI, browser extension, MCP server, and "pixel mode." The curl installer edits ~/.claude/settings.json and installs hooks and a statusline (it writes a .bak first, and INSTALL.md documents full uninstall).
Split licensing: the skill directory is MIT, but the compression engine, proxy runtime, and browse tooling are BSL-1.1 — free for first-party self-hosted use, commercial license required for hosted or embedded third-party services. GitHub reports the license as "Other."
The SKILL.md body itself costs roughly 1,069 tokens per load by the repo's own measurement — the tax its "pixel mode" exists to reduce.
The wenyan modes' advertised 80–90% reduction is in characters, not tokens; the README says so itself.
Auto-Clarity is an instruction to the model, not a hard mechanism — a model that ignores it can still compress a warning it shouldn't. 118 issues were open at verification time.

Facts come from the skill's own repository and docs on the date above, via the GitHub API — not from memory, and never from a sponsor.

Worth comparing

Claude Code output styles — The built-in, no-install way to make Claude Code terser: define a custom concise output style with /output-style. No levels, no wenyan, nothing to maintain.
amanattar/caveman-claude-skill — An early, simpler re-implementation of the same idea — 397 stars, untouched since April 2026. Only worth it if you want a single SKILL.md with no product suite attached.

← All MCP servers & skills