AI Pulse · Under the Radar

A developer published a tongue-in-cheek Claude Code skill that strips out conversational filler and makes the AI answer like a caveman. Blunt fragments, no niceties. The repository hit 96,000 GitHub stars in days, which is unusual velocity for a joke project, and two independent benchmarks measured 65 to 75 percent fewer output tokens with no loss in accuracy. The original repo is worth a look if you run Claude Code at any scale.

Why this matters: token economics are real money. If you’re calling Claude APIs dozens or hundreds of times a day, output tokens add up fast. A 70 percent reduction means your bill drops by the same ratio, assuming you’re paying per token. The benchmarks came from third parties, not the repo maintainer, so the claims have some credibility. No accuracy loss means the agent still solves the problem, it just skips the throat clearing.

This is the kind of thing we build into an AI command centre at /resources/omni, where you can test prompt variations, measure token spend per task, and roll out changes across your team without rewriting code. Most businesses don’t need caveman mode exactly, but the principle holds. Small prompt tweaks can cut costs without breaking output quality.

The broader point is that AI tooling is moving fast in weird directions. A joke repo becomes a real cost lever in 48 hours. If you’re not tracking these experiments, you miss opportunities to optimize the systems you already run. Watch the GitHub velocity on tools like this. When something hits four figures of stars per day, it usually means people found a real pain point, even if the solution started as a gag.