You’re simply placing your prompt in a structured format. It’s like completing a form. Instead of vague instructions, you provide precise, machine-readable input like this:<\/p>\n\n\n\n
{<\/mark><\/p>\n\n\n\n “task”: “summarize this article”,<\/mark><\/p>\n\n\n\n “audience”: “college students”,<\/mark><\/p>\n\n\n\n “length”: “100 words”,<\/mark><\/p>\n\n\n\n “tone”: “curious”<\/mark><\/p>\n\n\n\n }<\/mark><\/p>\n\n\n\n Not English. Not vibes. Just clear instructions.<\/p>\n\n\n\n LLMs (Large Language Models) don\u2019t “understand” language the way humans do. They follow patterns and structure. JSON is ultra-structured. It removes ambiguity. There\u2019s no guesswork involved; you’re telling them exactly what you want.<\/p>\n\n\n\n Can you write a tweet about dopamine detox?<\/p>\n\n\n\n {<\/mark><\/p>\n\n\n\n “task”: “write a tweet”,<\/mark><\/p>\n\n\n\n “topic”: “dopamine detox”,<\/mark><\/p>\n\n\n\n “style”: “viral”,<\/mark><\/p>\n\n\n\n “length”: “under 280 characters”<\/mark><\/p>\n\n\n\n }<\/mark><\/p>\n\n\n\n See the difference? Clear. Modular. Machine-readable.<\/p>\n\n\n\n You can nest<\/strong> the JSON:<\/p>\n\n\n\n {<\/mark><\/p>\n\n\n\n “task”: “write a thread”,<\/mark><\/p>\n\n\n\n “platform”: “twitter”,<\/mark><\/p>\n\n\n\n “structure”: {<\/mark><\/p>\n\n\n\n “hook”: “strong, short, curiosity-driven”,<\/mark><\/p>\n\n\n\n “body”: “3 core insights with examples”,<\/mark><\/p>\n\n\n\n “cta”: “ask a question to spark replies”<\/mark><\/p>\n\n\n\n },<\/mark><\/p>\n\n\n\n “topic”: “founder productivity systems”<\/mark><\/p>\n\n\n\n }<\/mark><\/p>\n\n\n\n You\u2019ve turned a messy prompt into clear, organized code.<\/p>\n\n\n\n 3 Basic Rules:<\/strong><\/p>\n\n\n\n {<\/mark><\/p>\n\n\n\n “task”: “generate a list”,<\/mark><\/p>\n\n\n\n “topic”: “books that improve thinking”,<\/mark><\/p>\n\n\n\n “audience”: “young entrepreneurs”,<\/mark><\/p>\n\n\n\n “output_format”: “markdown bullets”<\/mark><\/p>\n\n\n\n }<\/mark><\/p>\n\n\n\n Some even prefer<\/strong> it.<\/p>\n\n\n\n Recommend books that help me think clearer<\/p>\n\n\n\n {<\/mark><\/p>\n\n\n\n “task”: “recommend books”,<\/mark><\/p>\n\n\n\n “topic”: “thinking clearly”,<\/mark><\/p>\n\n\n\n “audience”: “entrepreneurs”,<\/mark><\/p>\n\n\n\n “output_format”: “list of 5 with one-sentence summaries”<\/mark><\/p>\n\n\n\n }<\/mark><\/p>\n\n\n\n Run both. The second one is sharper, easier to understand, and more on point.<\/p>\n\n\n\n Show me a product demo style video of a fitness app<\/p>\n\n\n\nWhy does this work so well?<\/h2>\n\n\n\n
Think of it like this:<\/h3>\n\n\n\n
Regular prompt:<\/strong><\/h3>\n\n\n\n
JSON-style prompt:<\/strong><\/h3>\n\n\n\n
Want even sharper outputs?<\/h3>\n\n\n\n
So how do you write JSON prompts?<\/h3>\n\n\n\n
\n
Example:<\/strong><\/h3>\n\n\n\n
This works across all major models:<\/h3>\n\n\n\n
\n
A quick comparison to prove it:<\/h3>\n\n\n\n
Normal prompt:<\/strong><\/h3>\n\n\n\n
JSON prompt:<\/strong><\/h3>\n\n\n\n
Here\u2019s one you can try in Veo 3 right now:<\/h2>\n\n\n\n
Simple prompt:<\/strong><\/h3>\n\n\n\n