Agreed on the judgment point. One place where structure helps: decomposing the prompt before you run it. Writing out role, objective, constraints, output format as separate fields forces you to answer "what do I actually want?" before the model does anything. You catch vague goals at prompt-writing time, not after you get a useless output.

That's the core idea behind flompt (https://github.com/Nyrok/flompt), a canvas where you build prompts from typed blocks and compile to XML. The act of filling in each block is itself a judgment exercise. Open-source if anyone wants to try it.