A developer dramatically cut the token usage of Anthropic's Claude by up to 75% by instructing the AI to communicate like a "caveman." This prompting technique, shared by a Reddit user , involves using short, 3-6 word sentences and eliminating all filler, directly addressing growing concerns over Claude's rising operational costs and recent usage limit reductions. This innovative approach surfaces as Anthropic faces user backlash for new usage caps and changes to its subscription model, making efficient token use more critical than ever. Users are hitting limits "way faster than expected," especially on paid accounts, according to BBC .
Optimizing Claude's Token Consumption
The developer's "caveman" prompting method relies on a few straightforward rules to minimize output length and, consequently, token consumption. These rules include using short sentences, typically between three and six words, and strictly avoiding any pleasantries or preambles. Additionally, the technique mandates running tools first, displaying results immediately, and then stopping without any additional narration. For instance, instead of "I will fix the code," the AI is instructed to say "Me fix code." This directness strips away unnecessary linguistic overhead.The impact of this method is significant. A typical web search task that might consume around 180 tokens with standard prompting can drop to approximately 45 tokens. This represents a 135-token saving per task , or a 75% reduction. Each "grunt swap"—replacing verbose AI output with a concise phrase—saves between 6 and 10 tokens. Over an entire task, this can lead to a 50-100 token saving, yielding a substantial 50-75% overall burn reduction when combined with a "tool-first" execution strategy.
Why Efficient Prompting Matters Now
The push for token efficiency comes amid significant shifts in Anthropic's service model. The company recently imposed stricter usage limits on its Claude chatbot, frustrating many devoted users. These changes have led to users, including those on paid subscriptions, encountering usage caps much more frequently than before, impacting their workflow, particularly for programming and coding tasks in Claude Code .Further complicating matters, Anthropic has ceased allowing its Claude subscriptions to integrate with third-party AI agent platforms like OpenClaw . This decision stems from the "outsized strain" such usage places on Anthropic's compute resources, according to Business Insider Africa. Users leveraging these platforms now must transition to a pay-as-you-go model or use a separate API key, which charges per token rather than offering the flat-rate access previously available with Pro and Max plans. This effectively translates to higher costs for heavy users, making every token count.






