Developer cuts AI coding token usage 70% by front-loading structured prompts

A software developer reduced Claude Code token consumption by 70% by shifting from conversational, on-the-fly prompting to writing detailed structured specs before each session. The core problem was that vague requests forced the AI to explore the codebase, infer intent, and guess at constraints — all inside the paid session, often repeatedly. The solution involved creating a 'master prompt' containing role context, a do-not-touch list, targeted discovery commands, numbered issues with expected outcomes, and clear acceptance criteria. Three supporting habits reinforced the approach: running codebase discovery before writing code, verifying results against real environments rather than assuming success, and limiting each task to a single concern per commit. The token savings came not from shorter prompts but from eliminating costly back-and-forth correction cycles that each carry the full conversation history as overhead.