The "think like an architect" framing resonates. One extension of this: the prompt itself benefits from the same decomposition discipline. Vague monolithic prompts sent to an AI coding tool tend to produce vague monolithic output — giving the AI a structured prompt with explicit role, objective, constraints, and output_format sections up front dramatically narrows the solution space.

I built flompt (https://flompt.dev) specifically for this — it lets you compose prompts from typed blocks visually before sending them, and has a decompose feature that breaks an existing unstructured prompt into those blocks. Makes the "design before you build" mindset apply to the prompt layer too.