When building applications that consume AI output programmatically, you need structured, parseable responses — not free-form text.

JSON Output Prompting

The most common structured format. Be explicit about the schema:

"Analyze this customer review and return a JSON object with exactly these keys: { "sentiment": "positive" | "negative" | "neutral", "topics": ["array of discussed topics"], "urgency": 1-5, "summary": "one sentence summary", "action_required": true | false } Return ONLY the JSON, no other text."

Tool Calling / Function Calling

Modern APIs (OpenAI, Anthropic, Google) support native function calling. Instead of asking the model to output JSON and parsing it yourself, you define a function schema and the model returns structured arguments. This is more reliable than raw JSON prompting.

XML for Complex Structures

For responses with nested, mixed-format content, XML tags work well:

"Return your analysis in this format: <analysis>

<summary>Brief overview</summary> <findings> <finding severity='high|medium|low'>Description</finding> </findings> <recommendation>Action items</recommendation> </analysis>"

Markdown Tables

For comparative data, request markdown tables:

"Compare these three products. Return a markdown table with columns: Feature, Product A, Product B, Product C"

Common Pitfalls

• Don't mix instructions with schema — Put the schema definition at the end, after all context
• Use examples — Show one complete example of the expected output format
• Handle edge cases — Tell the model what to output when data is missing ("use null for unknown fields")
• Validate outputs — Always validate the returned structure in your code; models can still produce malformed output

Production Pattern: Parse with Fallback

In production code, always implement a fallback:

1. Try to parse the structured output
2. If parsing fails, retry with a more explicit prompt
3. If retry fails, log the raw output and handle gracefully