A joke repo just made the best point about token economics
Caveman, a Claude Code skill that cuts about 65 percent of tokens by talking like a caveman, went viral. The punchline is a real budget line.
Caveman is a Claude Code skill that instructs the agent to communicate in caveman speak (“why use many token when few token do trick”) and claims to cut roughly 65 percent of tokens. It added over 2,800 GitHub stars in a single day. It is a joke, and it went viral because everyone running agents at scale immediately understood the point.
The point: tokens are now a cost center with the same shape as cloud spend in 2015. Nobody budgeted for it, everybody’s usage is growing, and the first finance review of an agentic program always contains a surprise. When a gag repo about terse prompting resonates this hard, it tells you how many teams are feeling the bill and how few have actual controls.
The serious version of this joke is a real discipline: context engineering and cost observability. Per-stage token budgets in agent pipelines, context assembly that retrieves only what the task needs, caching for repeated repository context, model routing that sends bulk work to cheap models, and cost-per-accepted-feature as a tracked metric. My rule from the enterprise side: a budget ceiling should be terminal for an agent run. The orchestrator may request more budget, but it cannot grant it to itself.
The first step, before any clever compression tricks: instrument cost per pipeline stage so you know where the tokens actually go. Most teams guess wrong. Then decide what the caveman version of each stage looks like.
Written by Adib Kadir. Product and engineering executive focused on rolling out AI at enterprise scale.
Start a conversation