A GitHub repository called 'caveman' — taglined '🪨 why use many token when few token do trick' — is circulating in prompt engineering communities for its claim that a Claude Code skill using highly compressed, grammar-stripped prompts can reduce token consumption by approximately 65%. The project describes itself as a technique that instructs Claude Code to communicate in minimal, caveman-style language during extended coding sessions.
The core premise is straightforward: in agentic coding workflows where Claude Code may exchange dozens or hundreds of messages with itself or with scaffolding code, verbose natural language in system prompts and responses consumes tokens that translate directly into API costs and latency. By establishing a communication convention that strips syntax to bare essentials, the author argues that semantic intent is preserved while token overhead is dramatically reduced.
The 65% figure has not been independently verified in peer-reviewed conditions, and results are likely to vary depending on task type, model version, and workflow structure. However, the concept resonates with a real problem in production agentic systems: token costs compound quickly across long-running agent loops, and any systematic reduction in prompt verbosity can have meaningful economic impact at scale.
The project sits at the intersection of two active areas of developer interest — cost optimization for AI agents and prompt engineering as a disciplined craft. Whether or not the specific numbers hold under scrutiny, the underlying approach of designing communication protocols between AI components for efficiency rather than human readability represents an important design principle for the next generation of agentic infrastructure.