caveman: why use many token when few token do trick

What It Is

caveman is a Claude Code skill — also compatible with Cursor, Windsurf, Cline, and 30+ other agents — that compresses your AI output by instructing it to communicate like a caveman: no pleasantries, no restating the question, no three-paragraph intro before the code block. Technical content (code, commands, URLs) stays byte-for-byte intact. Everything else gets squeezed.

The headline number: 65% average output token reduction across 10 benchmark tasks, ranging from 22–87% depending on how verbose the baseline would have been. Postgres pool setup: 2,347 tokens down to 380. Auth middleware debug walkthrough: 704 to 121. React re-render explanation: 1,180 to 159.

When It Helps (And When It Does Not)

If you are paying per output token, caveman pays off fast on verbose tasks — architecture explanations, step-by-step debugging walkthroughs, anything where Claude's default is "let me tell you everything I know about this topic first." Six compression levels (lite, full, ultra, and a wenyan mode that encodes responses as classical Chinese for maximum compression) let you dial it in. /caveman-commit generates conventional commits under 50 chars; /caveman-compress <file> rewrites memory/context files for a separate ~46% input reduction.

The real caveat, which the project calls out honestly: caveman only shrinks output tokens. Input and reasoning tokens are untouched. The skill file itself adds ~1–1.5k input tokens per turn, so whole-session savings run meaningfully smaller than the 65% headline — and on already-terse workloads they can go net-negative. The /caveman-stats command logs per-session and lifetime savings so you can check whether it is actually paying off for your workload. A March 2026 paper found constraining models to brief answers improved accuracy ~26 points on some benchmarks, which is a nice bonus if it holds for your use case — but the primary pitch is cost, not quality.

Install

# macOS / Linux / WSL
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

After install, /caveman toggles compression on for the current session. Re-run at any level: /caveman lite, /caveman ultra, /caveman wenyan.

Credit

Built by Julius Brussee — JuliusBrussee/caveman, MIT licensed, ~90k stars. Part of a small ecosystem of sibling skills (cavemem, cavekit, caveman-code) if you want to go deeper.