Why use many token when few do trick

A practical tip for reducing token usage and costs when using Claude Code or similar AI coding assistants. The 'Caveman' prompting style instructs the AI to drop pleasantries, articles, hedging language, and filler phrases, communicating in minimal but technically precise language. Examples show token reductions from 69 to 19 tokens and 1,180 to 159 tokens (87% savings). The approach is backed by a March 2026 study suggesting brevity constraints improve LLM accuracy by 26 percentage points. The Caveman spec is available on GitHub and supports multiple levels: light trimming, standard, and ultra-maximum abbreviation.