Select a prompt and refine it with an LLM. Two modes: template-driven structured optimization, and template-free light polishing (disambiguate / dedupe / mark missing info with placeholders).