AI Pulse · Under the Radar

A developer named Julius Brussee built a Claude prompt that tells the AI to talk like a caveman, and it cuts token usage by roughly 65%. The trick is simple: strip out polite filler (“I understand”, “Let me help you with that”) but keep code blocks, commands, and technical output untouched. One example in the repo shows a bug explanation shrinking from 1,180 tokens down to 159.

The GitHub project hit 93,500 stars and is climbing at about 817 stars a day, which means a lot of people are tired of paying for AI politeness. Brussee has already spun out a small ecosystem around it: cavemem for memory management, cavekit for developer tools, and a waitlist for v2 that adds team dashboards and analytics.

Why it matters

Token costs add up fast when you run agents at scale. If your assistant writes three paragraphs of preamble before every answer, you pay for every word. Cutting that overhead by two thirds without losing the actual work means cheaper API bills and faster responses. It also matters for context windows: the less space wasted on fluff, the more room for real data.

This is exactly the kind of efficiency layer we bake into the Omni Command Centre, where every token saved across hundreds of tasks compounds into real savings. You do not need to rewrite prompts by hand. You need systems that route tasks to the right model, compress what can be compressed, and log what actually costs you money. Caveman speak is one tactic. The broader principle is treating tokens like the budget line item they are.