Structured Outputs

Structured outputs ensure AI models return data in a specific, parseable format. JSON mode and schema enforcement eliminate parsing errors and make AI responses directly usable in applications.

Basic JSON Mode

Simple Object:

Return the following information as JSON:
{
  "name": "string",
  "age": "number",
  "city": "string"
}

Text: "John is 32 years old and lives in Seattle."

Response:

{
  "name": "John",
  "age": 32,
  "city": "Seattle"
}

Schema Definition

With JSON Schema:

{
  "type": "object",
  "properties": {
    "summary": {
      "type": "string",
      "description": "Brief summary in 1-2 sentences"
    },
    "sentiment": {
      "type": "string",
      "enum": ["positive", "negative", "neutral"]
    },
    "key_topics": {
      "type": "array",
      "items": {"type": "string"},
      "maxItems": 5
    }
  },
  "required": ["summary", "sentiment", "key_topics"]
}

Extraction Patterns

Entity Extraction:

Extract entities from the text as JSON:

{
  "people": [{"name": "string", "role": "string"}],
  "organizations": ["string"],
  "locations": ["string"],
  "dates": ["string"],
  "monetary_values": [{"amount": "number", "currency": "string"}]
}

Text: {input_text}

Data Transformation:

Convert the following unstructured text to structured data:

Input format: Free text
Output format:
{
  "records": [
    {
      "field1": "value",
      "field2": "value",
      "field3": "value"
    }
  ],
  "total_count": "number",
  "extraction_confidence": "number (0-1)"
}

Validation Techniques

Required Fields:

Return JSON with these REQUIRED fields:
- title (string)
- description (string)
- priority (number, 1-5)

Missing any required field = invalid response.

Enum Constraints:

Status must be one of: ["pending", "active", "completed", "cancelled"]
Priority must be one of: ["low", "medium", "high", "critical"]
Category must match: ["bug", "feature", "improvement", "documentation"]

Format Validation:

Dates in ISO 8601 format: "2025-01-15T10:30:00Z"
Emails must match: [email protected]
URLs must start with: https://

Common Patterns

Pattern	Use Case
Single object	Extract one entity
Array of objects	Parse multiple items
Nested objects	Complex hierarchical data
Union types	Variable response structures

Error Handling

If information is missing or unclear:
{
  "extracted_data": {...},
  "missing_fields": ["field1", "field2"],
  "confidence": 0.85,
  "notes": "Explanation of uncertainties"
}

Tips

Be explicit about types - Specify string vs number vs boolean
Use enums for fixed options - Reduces invalid values
Include descriptions - Help model understand field purposes
Set constraints - Min/max values, string lengths, array sizes
Handle missing data - Define how to represent unknowns

Needle-in-Haystack: Finding Specifics in Massive Claude Contexts

Prompt patterns for targeted information retrieval from 200K token contexts. Multi-hop question answering, verification strategies, and techniques to ensure Claude finds what you're looking for in massive documents.

DeepSeek Tool Calls with Thinking: reasoning_content Management

Combine DeepSeek's tool calls with thinking mode. The mandatory reasoning_content passback in tool loops, strict mode for JSON schema enforcement, and error handling patterns to avoid 400 errors.

CoT vs Extended Thinking: When to Use Which

Compare chain-of-thought prompting vs Claude's extended thinking. Understand performance differences, use cases for each approach, and hybrid strategies that combine both for optimal results.

Structured Outputs (JSON Mode)