High Token Usage in AI Tools Signals Inefficiency, Not Productivity

A software developer argues that high token consumption in AI-assisted coding does not correlate with better business outcomes or more shipped work. Drawing an analogy to lighting a room, the author contends that unfocused, sprawling prompts waste computational resources without producing usable results. The author identified recurring inefficiency patterns and developed personal strategies — such as pre-reading the codebase, compressing prompts, and specifying minimal viable solutions — to reduce token waste. From a business standpoint, the relevant metric is not tokens used or code generated, but what actually shipped and whether it justified the cost. The author concludes that token limits, rather than being obstacles, impose a useful discipline that unlimited capacity would eliminate.