Caveman Technique: Revolutionary AI Approach Lowers Token Costs Up to 75%

The discovery that it is possible to have Claude — an artificial intelligence (AI) — communicate in a 'caveman style' has captured the attention of the tech community. This approach limits output to short, powerful sentences, leading to a reduction in the cost of output tokens by as much as 75%. For developers who rely on cost-saving strategies, this insight is surprisingly valuable.

What makes the method so intriguing is the way it streamlines communication. Unlike the usual, narrative style typically employed by AI, the caveman approach steers clear of verbose language and unnecessary explanations. This results in a drastic reduction in the number of tokens required for a standard task. A typical web search query, which normally requires around 180 tokens, can be reduced to approximately 45 tokens. This has concrete implications for developers, where the token savings can significantly lower operational costs in daily use.

However, an important nuance must be made. The method does not touch the input context—that is, the entire call history, attachments, and system instructions that the model rereads with every interaction. This input often represents a larger amount of data than the output, especially in more complex coding sessions. In reality, developers can still achieve significant savings, but the total cost reduction is typically around 25%, rather than the previously mentioned 75%. This is still a substantial reduction that is of interest to investors and analysts, particularly in the context of emerging cost models.

It is also crucial to consider how the limitations of this communication style might affect the intelligence of the AI. While some research findings suggest that these simplified conversations may impair the model's other reasoning skills, the question of how these cognitive limitations influence the output remains unclear. It is a lesson for the future of AI development: how simpler communication does not always guarantee a better result.

The technique has since found its way to GitHub, where developer Shawnchee has compiled the rules into a standalone caveman skill. This skill is compatible not only with Claude but also with other AI agents such as Cursor and Copilot. The approach has essentially been reduced to ten concise rules that disrupt the process: no filler sentences, execution first, then explanation, and errors are simply addressed rather than reported. Such benchmarks have reported that output reductions of up to 68% are possible on web searches.

In addition, developer Julius Brussee has presented an alternative approach that further expands on the idea of ​​caveman communication. This model, presented in a SKILL.md file, emphasizes the importance of technical precision without unnecessary distractions. The results show a promising trend in which short, efficient communication not only saves costs but also increases productivity.

In the broader context of the AI ​​cost structure, this caveman technique offers a sharper view of the current market. Anthropic, the developer behind Claude, is known for its high prices per output token. For developers using agentic workflows with a large number of interactions per session, this output representation is more than an aesthetic consideration; it is a significant cost. The choice of a simple style can then be translated into financial terms, where every saved token contributes to overall profitability.

The caveman skill can be easily installed with a single command via skills.sh and can be applied globally to various projects. Whether this approach actually makes Claude less talkative is secondary to the fact that it also significantly alleviates developer frustration.

Frequently Asked Questions

How much token saving can the caveman technique actually yield?
The technique can reduce the output tokens by an average of 25% to 75%, depending on the task and the input context.

What are the potential consequences for AI performance when using this method?
There are concerns that shifting to a simpler communication style may affect the reasoning skills of the AI, which means that a trade-off must be made between efficiency and quality.

How can developers implement the caveman skill in their projects?
The caveman skill can be easily installed with a single command via skills.sh and is then available for use in various projects and workflows.