This Developer Made Claude Code Talk Like a Caveman to Cut Token Costs in Half

A developer named Julius Brussee published a tiny GitHub repo called "caveman" that does one thing: it forces Claude Code to respond in terse caveman language. No full sentences. No explanations. Just the minimum words needed to communicate what's happening.

The repo hit the front page of Hacker News with 373 points. The reactions ranged from "this is ridiculous" to "this is genuinely useful" - and the numbers back up the useful camp.

What the caveman skill actually does

Claude Code has a feature called "skills" - custom instructions that modify how Claude behaves during a coding session. Skills are defined in a CLAUDE.md file and loaded automatically when you open a project. They tell Claude things like: always write tests before code, prefer functional patterns, never use var.

The caveman skill adds a simple instruction: respond like a caveman. Short sentences. Drop articles and prepositions. Grunt instead of explain. "Claude find bug. Here fix." instead of "I've identified the issue in your authentication middleware and here's a comprehensive explanation of what's happening and how to fix it..."

The result is Claude Code that communicates like a terse senior engineer instead of a verbose assistant trying to demonstrate thoroughness.

Why this cuts costs

When you use Claude Code, you pay for every token - both what you send to the model and what comes back. Claude's output is often the larger cost driver because the model tends to explain, qualify, and elaborate by default.

A typical Claude response to "what's wrong with this function" might run 400-600 tokens of explanation before giving you the fix. A caveman response gives you 50-100 tokens of signal: what's wrong, what to change, done.

For developers running heavy Claude Code sessions - multiple hours of autonomous task execution, code review, debugging - the output token savings add up quickly. Users reported 40-50% reductions in token usage for sessions that involve a lot of back-and-forth with Claude.

The tradeoff

The caveman mode is not suitable for every situation. When you need Claude to reason carefully through a complex problem - debugging a race condition, designing an architecture, reviewing security implications - you want the full explanation. Cutting the output to caveman levels means Claude is also cutting the chain-of-thought reasoning that helps it get hard problems right.

Where it works well: status updates, simple fixes, linting errors, routine refactors - anything where you already understand the problem and just need Claude to do the work.

The smarter use of this skill is context-dependent: use it for sessions where you're in execution mode, turn it off when you need Claude to think out loud.

The bigger point about token cost

The caveman skill is a fun example of a real problem: Claude Code can cost serious money for heavy users, and most of that cost is output tokens from responses that are more verbose than they need to be.

Other techniques that reduce token costs: keeping CLAUDE.md instructions tight and specific (every instruction is re-sent as context in every request), using the escape hatch to stop Claude before a long explanation, and breaking large tasks into smaller sessions rather than keeping long context windows open.

The caveman skill is the bluntest version of this - just telling Claude to shut up and code. But the instinct behind it is sound. AI assistants default to verbose because verbosity usually reads as thoroughness. For experienced developers who can evaluate output quality without needing every step explained, that verbosity is pure cost.

The repo is at github.com/JuliusBrussee/caveman if you want to try it. Add it to your CLAUDE.md and run a session. The savings are real. The vibe is genuinely weird.