TOOLS & SKILLS
Caveman Prompt Compression: Slash Claude Token Usage by 65% (Free Skill) 🛠️
A free GitHub skill cuts your Claude token use by 75% — just by making it talk like a caveman. Same answer. A quarter of the words. None of the burnt credits.
The Problem
Opus 4.7 is the smartest model Claude has ever shipped. It’s also burning through usage limits in a way 4.6 never did. Long, polished, over-explained answers eat tokens fast.
A developer named Julius Brussee built a free GitHub skill called caveman with the tagline “why use many token when few token do trick.” It rewrites Claude’s responses in caveman language. Same useful answer, way fewer words.
What It Looks Like
Before
“Certainly! I’d be happy to help you with that. The next step you’ll want to take is to navigate over to your settings menu and toggle the option on.”
After (Full Mode)
“Next step. Open settings. Toggle on.”
Same answer. About a quarter of the tokens. Across hundreds of conversations, that adds up to entire days of extra usage.
3 Intensity Levels
Lite
Keeps full sentences and articles, just removes filler and corporate-speak. Easiest to read, smallest token savings. Good if you’re sharing Claude’s output with other people.
Full — The Sweet Spot
Drops articles, uses short fragments, swaps in shorter synonyms. ~75% token reduction. Still completely readable. This is the one most people should use.
Ultra
Maximum compression. Abbreviates common terms (DB, auth, config, fn), strips conjunctions, uses arrows for cause-and-effect (X → Y). Highest savings, takes a second to read. Use when you’re grinding through high-volume work.
2 Ways Depending on Your Setup
If You Use Claude Code
One command in your terminal:
npx skills add JuliusBrussee/caveman
Done. Caveman activates automatically. Toggle modes with /caveman full.
If You Use Claude.ai
This is one skill. The Weekend Bootcamp shows you how to set up Claude as a complete operating system for your specific job — Projects, Skills, connectors, scheduled tasks — all wired together in one weekend.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.