AI Pulse · Under the Radar

A GitHub repository called Caveman hit 94,000 stars in under four months, one of the fastest climbs in recent memory at roughly 813 stars a day. It’s a custom skill for Claude that forces the model to respond in terse, caveman-style fragments instead of verbose sentences. The point is simple: cut token spend.

The repo’s self-published benchmarks claim token reductions between 22% and 87% across ten test prompts. Those numbers aren’t independently verified, but the approach is straightforward enough that anyone can test it. You tell Claude to strip out articles, conjunctions, and filler words. The output reads blunt and choppy, but if you’re running hundreds of agent calls a day, the cost savings add up fast.

This matters if you’re managing AI spend at scale. Most companies don’t realize how much they’re paying for polite phrasing and conversational padding in their agent responses. A tool like Caveman won’t work everywhere, internal customer-facing outputs still need tone, but for backend tasks like data extraction, classification, or internal summaries, terse is fine. The trick is knowing when to apply it and tracking whether the trade-off actually saves you money without breaking downstream workflows. That kind of cost discipline is exactly what we build into systems like the Omni Command Centre, where you can test different prompt strategies and measure token burn in real time.

The repo’s velocity suggests developers are hungry for practical cost controls, not just model performance. If you’re running agents in production, this is worth a quick test. Clone the skill, run it against a few of your own prompts, and see if the output still works for your use case. If it does, you just found a line item to trim.