AI Pulse · Under the Radar

A Claude Code skill called caveman is tearing across GitHub at 782 stars a day, one of the fastest climbs this cycle. It does one thing: strips filler language from agent output to cut token use. No “I appreciate your patience”, no “let me help you with that”, no hedging. Just the answer.

The repo hit 95.7k stars in days, and the reason it resonates is simple. Every token an agent generates costs money and time. When you multiply that across hundreds or thousands of queries, the waste compounds fast. Caveman forces the model to answer like a terse engineer, not a customer service script. The result is cheaper, faster, and often clearer.

What makes this interesting is that it’s the same posture we already enforce in EDNA’s house style, independent convergence. We didn’t invent brevity, but we’ve been training models to skip the fluff for months because it matters in production. When you’re running an AI command centre that handles real workflow, every wasted token is a drag on throughput and cost. The caveman repo proves the broader market is catching up to what tight operators already knew.

Why this matters

If you’re building AI into your business, you want output that gets to the point. The novelty here isn’t the technique, it’s the velocity of adoption. Developers are tired of paying for politeness. This is the kind of efficiency tweak we bake into the Omni Command Centre by default, because trimming overhead at the prompt level compounds across every interaction. You don’t need a viral GitHub repo to do it, you just need to treat tokens like the cost line item they are.