AI Pulse · Under the Radar

A GitHub repo called “Caveman” started as a joke and is now one of the fastest-growing projects on the platform. It’s a skill for Claude Code that strips AI agent responses down to terse, blunt fragments. Think “file created” instead of “I’ve successfully created the file for you and placed it in the appropriate directory.” The repo has pulled 93,172 stars in 113 days and is still adding roughly 822 new stars every day.

The tool offers six intensity levels. At the low end, responses stay relatively normal. At the high end, there’s a “wenyan” mode that compresses output into classical Chinese brevity. Independent reviews confirm the headline claim: output tokens drop by about 65%. That matters because output tokens cost more than input tokens on most API pricing tiers, and verbose agent chatter adds up fast when you’re running hundreds or thousands of tasks.

The catch is that the skill itself adds between 1,000 and 1,500 input tokens per turn, according to the original repo. So you’re trading input cost for output savings. Whether that trade makes sense depends on your usage pattern. If your agents generate long responses, the math works. If they’re already concise, you might lose money.

What’s interesting here isn’t the tool itself. It’s the signal. Developers are tired of chatty agents and willing to adopt a half-serious hack to fix it. That tells you something about where friction lives in production AI workflows. When we build systems like the Omni Command Centre, controlling verbosity and token spend is baked into the orchestration layer, not bolted on afterward. You shouldn’t need a meme repo to make your agents efficient. But right now, plenty of teams do.