How to format LLM outputs for visual tools, design reliable handoff payloads, and test multi-modal AI pipelines where text rubrics miss image quality.