What is a JSON prompt?
A JSON prompt organizes instructions as JSON, asks an AI service to return JSON, or does both. JSON stores information in named fields and values. It is useful when you want results that can be checked, compared or passed into another step of a workflow.
You can write the instructions in ordinary language and request a JSON response. For example, “Summarize these reviews and return JSON with sentiment, themes and actions” is a plain-language instruction asking for structured output.
Choose the structure before writing the prompt
Start with the fields you actually need. For a content idea, that might be platform, topic, caption and call to action. For customer feedback, it might be a summary, a list of recurring issues and the next action. Keep field names consistent so the next person or tool can understand the result.
Specify how to handle missing information. A helpful instruction is: “Use null when a requested fact is not supplied. Do not invent a value.” This makes gaps visible rather than turning them into confident-looking guesses.
A complete example
Here is a prompt for a fictional neighborhood bakery. The only supplied facts are the business type and a Saturday bread tasting.
Create one Instagram post idea for a fictional neighborhood bakery promoting a Saturday bread tasting.
Use only the supplied facts. Do not invent prices, opening hours or an address.
Return only a JSON object with these string fields:
platform, topic, caption, call_to_action.Illustrative response
{
"platform": "Instagram",
"topic": "Saturday bread tasting",
"caption": "Discover your next favorite loaf at our Saturday bread tasting.",
"call_to_action": "Join us for the tasting."
}This is a written example of the requested structure. It is not a measured result from an AI-service comparison.
Check both the format and the content
First, confirm that the response parses as JSON: property names and strings need double quotes, and trailing commas are not valid. Next, check that the expected fields exist and contain the right kind of value. Finally, review the facts and usefulness of the text inside those fields.
A correctly formatted object can still contain an inaccurate caption, an unsupported claim or the wrong recommendation. Formatting and factual accuracy are separate checks.
When a chat prompt is enough—and when to use an API feature
For a draft you will review manually, requesting JSON in a chat can be a useful starting point. For software that depends on a specific response shape, consider the AI provider’s structured-output features and validate the result in your application. OpenAI and Gemini document schema-based structured outputs; a sentence requesting JSON does not provide the same guarantees.
Customize a template in five steps
- Choose a prompt that matches your task.
- Replace the bracketed inputs with your source material and business details.
- Keep only the fields your workflow needs.
- Explain how to handle unknown facts, empty results and exceptions.
- Review the response before publishing it or passing it to another tool.
Explore our JSON prompt directory for starting templates, or try the Customer Review Analyzer for a feedback task.