Output Formatting & JSON Mode
The Waiter Who Fills in the Form
6 min read
The Waiter Who Fills in the Form
Instead of describing the order in a paragraph, the waiter fills in structured slots — table number, dish, quantity, special requests.
An AI that only produces prose is like a waiter who describes every order in a long essay. Useful for reading, useless for the kitchen computer system. JSON mode forces the AI to return structured, machine-readable data — the bridge between conversational AI and real software applications. This is what makes AI outputs actually programmable.
In Plain English
JSON mode (or structured output) tells the AI to return data in a fixed format — like JSON — instead of freeform text. This makes AI outputs directly usable by your application code without needing to parse messy prose.
The Technical Picture
JSON mode constrains the model's output distribution to valid JSON using constrained decoding or fine-tuning. The OpenAI and Anthropic APIs support structured outputs via response_format parameters, enabling schema-validated JSON generation for reliable programmatic consumption.
Real-World Examples
- Extracting structured product data from unstructured reviews
- Lovable uses structured prompts to get consistent component code from Claude
- Generating a list of tasks as JSON array instead of a bullet-point paragraph
JSON mode is what turns an AI chatbot into a programmable API — the bridge from language to data.