Structured Output: Get JSON and More from LLMs

Structured output is the practice of getting an AI model to respond in a specific, predictable format — JSON, XML, markdown tables, CSV, or any schema you define. Instead of parsing free-text prose, you get data you can plug directly into code, databases, or downstream systems.

Why It Matters

A model’s default response is natural language — great for conversation, terrible for automation. If you’re building anything that processes AI output programmatically, you need structure. Extracting entities from emails? You want JSON. Generating a comparison? You want a table. Feeding results into another system? You need a consistent schema.

How to Get Structured Output

The simplest approach: show the format you want. Models are excellent at pattern-matching, so an example does more than a paragraph of instructions.

Extract the name, role, and company from this email signature.
Return the result as JSON:

{"name": "...", "role": "...", "company": "..."}

Signature: Jane Park, VP Engineering, Acme Corp

For more complex schemas, add inline constraints to guide each field:

Analyze this product review. Return your analysis as JSON:

{
  "sentiment": "positive" | "negative" | "mixed",
  "key_topics": ["topic1", "topic2"],
  "confidence": 0.0-1.0
}

Structuring Your Prompts Too

Structure isn’t just for output — it improves your inputs. Using delimiters like XML tags to organize your prompt helps the model distinguish instructions from context:

<instructions>
Summarize the document below in three bullet points.
</instructions>
<document>
{{DOCUMENT_TEXT}}
</document>

This is especially powerful when prompts grow complex. Clear boundaries reduce misinterpretation and let you swap components without rewriting the whole prompt.

Tips

• Provide an example of the exact output format — this is the single most effective technique
• Name fields clearly — sentiment_score beats ss
• Specify constraints inline — "rating": 1-5 or "status": "approved" | "rejected"
• Use API features when available — both Claude and GPT offer structured output modes that enforce a schema at the API level, guaranteeing valid output

When your output is structured, everything downstream gets simpler — parsing, validation, logging, and debugging. Now that you can control what comes back, the next step is building prompts you can reuse across many different inputs.