prompting in json or xml format increases LLM output by 10x why because structured formats give the model clear boundaries and expectations when you ask for unstructured text, the LLM has to guess: how long should this be? what sections do I need? when am I done? what level of detail? but with JSON/XML you're literally providing a template: json{ "summary": "", "key_points": [], "analysis": "", "recommendations": [] } now the model knows exactly what to fill and roughly how much content each section needs it's like the difference between "write something about cars" vs "fill out this car review form with these 12 specific fields" the structure removes decision paralysis and gives the model permission to be comprehensive plus JSON/XML naturally encourages the model to think in organized chunks rather than just streaming text until it feels "done" [here's the meta hack] don't even try to write structured prompts from scratch instead: 1. voice-to-text brain dump everything you want analyzed/researched/written about 2. paste that messy transcript to the AI 3. ask it to "create a JSON prompt structure based on this brain dump that would get the best possible output" 4. take that generated JSON template 5. send it back as your actual prompt you get all the benefits of structured prompting without having to think through the structure yourself the AI knows what fields would be most useful better than you do anyway turns your stream of consciousness into a professional-grade prompt in 30 seconds try it yourself - ask for the same analysis in paragraph form vs structured format the structured version will be 3-5x longer and way more detailed every time