Someone discovered that if you tell Claude to drop the pleasantries — no "Certainly! I'd be happy to help with that" preambles, no hedging, no filler — the answers get shorter, cheaper, and sometimes better.
The project is called Caveman. It's a small add-on for Claude Code (Anthropic's AI coding assistant) that instructs the AI to respond in blunt, stripped-down language. Think grunts and facts, nothing more. It racked up 15,000 admirers on GitHub within days.
If you or your team use Claude heavily — writing proposals, summarising research, handling customer queries — you're paying per word the AI produces. Caveman's own measurements show savings between 22% and 87% per conversation. A research paper from early 2026 even found that forcing brevity can make the AI more accurate, not less.
Less fluff. Lower bill. Sharper answers. That's the whole story.
Most AI tools are tuned to sound helpful and polished. That's often unnecessary padding. If your use case is internal — drafting, summarising, researching — you probably don't need the AI to be charming. You just need it to be right.
Ask your next Claude session to "be brief, skip pleasantries." You might be surprised how much quieter — and more useful — it becomes.
Token — The unit AI companies use to measure text. Roughly one word, or part of a word. You pay per token.
Claude Code — Anthropic's AI assistant designed to help with technical and writing tasks from a command line (a text-based interface on your computer).
API costs — What you pay to use an AI service programmatically, usually based on how much text goes in and comes out.
GitHub — A website where developers share open-source projects. Think of it as a public library of tools anyone can pick up and use.