Claude Code: How to Reduce Your Token Usage by 75% Without Losing Precision

AI-assisted development (Claude Code, Cursor, Windsurf) has radically changed our productivity. However, a "hidden cost" weighs on every session: verbosity. Current agents are programmed to be polite and explanatory, which needlessly consumes a massive amount of tokens.

To address this issue, an open-source tool called Caveman proposes a radical approach: forcing the AI to adopt a telegraphic style. The result? Responses that are 3x faster and a token bill divided by four.

Project link: GitHub - JuliusBrussee/caveman

The Problem: The "Verbosity Tax"

Every time Claude Code generates a sentence like "I would be happy to help you fix this bug, here is a detailed analysis of your React component...", you pay for text that brings zero technical value.

This verbosity poses three major problems:

• Latency: The more the AI writes, the longer you wait in front of your terminal.
• Cost: Output tokens are the most expensive. On large projects, this "politeness" can represent hundreds of dollars a month.
• Cognitive noise: The developer must sort through the information to find the relevant code hidden within paragraphs of text.

The Solution: The "Caveman" Approach

Caveman is an extension (skill/plugin) designed for Claude Code, Cursor, and other agents. It injects system instructions that force the AI to remove all the fluff (articles, politeness, complex grammatical structures) to keep only the raw technical substance.

Efficiency Comparison

The Science Behind Brevity: Fewer Words = More Precision?

One might think that limiting the AI's expression reduces its reasoning capabilities. It's actually the opposite.

A study published in March 2026 ("Brevity Constraints Reverse Performance Hierarchies in Language Models") shows that constraining a model to be brief increases its precision by nearly 26% on certain benchmarks. By eliminating filler, the AI remains focused on pure logic and reduces the risk of verbal hallucinations.

Important Note: Caveman does not reduce the "thinking tokens" of models like Claude 3.7 Sonnet. The AI still thinks just as much, it simply communicates its conclusions in an optimized way.

Beyond Output: Optimizing Input with caveman-compress

The Caveman ecosystem is not limited to text generation. It includes a powerful utility to compress your own context files (like your project documentation or your CLAUDE.md).

Using caveman-compress, you can reduce the size of your reference files by an average of 45%. The tool creates a compressed version that the agent reads at each session, thus saving thousands of input tokens while retaining all technical meaning.

Installation and Configuration

Installation is fast and compatible with most development environments.

For Claude Code

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

For Cursor, Windsurf, or GitHub Copilot

You can add the rules via npx:

npx skills add JuliusBrussee/caveman

Once installed, you can change the intensity level according to your needs:

• /caveman lite: Professional, no frills.
• /caveman full: Full caveman mode (no more articles).
• /caveman ultra: Maximum compression (extreme telegraphic style).

Conclusion

For any developer intensively using AI agents, Caveman is an indispensable optimization. By treating tokens as a precious resource, you gain in speed, clarity, and budget.

Find the complete project and contribute on GitHub: 👉 https://github.com/JuliusBrussee/caveman