Structured Outputs

Why structured outputs

Most people get free-form text from AI and then manually reformat it. That's a waste of time. When you tell AI exactly what format you want, you get output you can use directly — paste into a spreadsheet, process with a tool, or share with colleagues without editing.

Structured outputs are especially valuable in three situations: you're processing output programmatically (JSON, CSV), you need consistent format for repeated use (templates), or you want output that drops straight into another document (tables, lists).

Markdown tables

The simplest structured format. Just say 'Format the output as a table' and most models will create a markdown table. But for consistent results, be more specific — define the exact columns.

Prompt: 'Compare the 3 most popular task management tools (Asana, Monday.com, Trello). Output as a table with columns: Name, Monthly price (team of 10), Main advantage, Main disadvantage, Ideal for.' — This tells the AI exactly which columns you want and in what order. The result will be a consistent table you can paste directly into a document or presentation.

Structured bullet lists

When you need a structured list, define the format of each item. Instead of 'Write a list of tips' try: 'Write 5 tips. Format each tip as: tip name — 1 sentence explanation. Example: [concrete usage example].' This gives you a consistent format for every item.

JSON outputs

JSON is ideal when you need to process data with a program, import into a tool, or simply want perfectly structured output. The key: define the schema upfront and give an example with specific value types.

// Prompt: Analyze this text and return the result as JSON:
{
  "sentiment": "positive | negative | neutral",
  "confidence": 0.0-1.0,
  "key_topics": ["topic1", "topic2"],
  "summary": "1 sentence summary",
  "action_items": [
    {
      "task": "description",
      "priority": "high | medium | low"
    }
  ]
}

When you give AI a specific JSON schema with example value types (string, number, array), the model follows it precisely. If you need guaranteed valid JSON (for programmatic processing), the OpenAI API offers 'JSON mode' and 'Structured Outputs' — the model is forced to return valid JSON.

Schema-driven prompting

The most reliable technique for structured outputs: give AI an example of the complete structure with labels instead of values. AI then fills in real data into your structure. This works because you're telling the model exactly what shape the output should have — the model then predicts tokens to match your template.

XML and tagged context

For complex prompts with multiple inputs, use XML-like tags to distinguish different parts. Claude explicitly recommends this approach. Example: '<document>document text</document> <instructions>your instructions</instructions> <output_format>desired format</output_format>'. Tags help the model unambiguously distinguish what is context, what is instruction, and what is format.

Consistency and common mistakes

When processing multiple items the same way, define the format once and repeat it. AI adding text around the structure? Add: 'Return ONLY [format], no explanation before or after.' AI changing format between items? Add a sample example of the first item. AI returning invalid JSON? Simplify the schema — fewer nested objects, shorter values, fewer optional fields.