Julius Brussee · 2026-04-08 · notable
caveman — Claude Code skill that compresses chat output
Claude Code skill that strips filler from an agent's chat output while leaving code untouched — Brussee reports 22-87% token savings, averaging about 65%.
'Why use many token when few token do trick' — a Claude Code skill that drops articles and filler to cut about 65 percent of an agent's chat tokens.
Key specs
| GitHub stars | 17.6k |
|---|---|
| Avg token savings | 65% |
| Range | 22-87% |
What is it?
Caveman is a small skill / plugin for Claude Code (and Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and 40+ others) that makes the agent talk in compressed 'caveman' English in its non-code output. It drops articles, pleasantries and hedges, but leaves code blocks, commits and PR descriptions normal.
How does it work?
The skill is a system-prompt rule applied through the Claude Code skill mechanism. The agent's reasoning and chat output is constrained to a terse style; technical content (code, JSON, tests) is exempted. There are multiple intensity levels — Lite, Full, Ultra and a Classical Chinese mode — for trading brevity against politeness. Brussee benchmarks token savings between 22% and 87% across coding tasks, averaging about 65%.
Why does it matter?
For anyone running Claude Code or Codex on long agentic workflows, a 65 percent cut on the chat layer is a direct bill reduction with no quality loss on code. It is also the most fun-to-read example of how the new Claude Code skills mechanism is going to reshape CLI-based agents.
Who is it for?
Heavy Claude Code and Codex users watching their token bill.
Try it
claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman