Skip to main content

Output Formatting

Get structured, parsable, and immediately usable output from LLMs

Make Output Machine-Readable

The difference between usable and unusable AI output often comes down to formatting. Structured output enables automation, integration, and programmatic processing of LLM responses.

JSON & XML Structured Outputs

1

Specify the Exact Schema

Don't just ask for JSON — provide the exact structure with field names, types, and nesting.

Extract contact information from this email and return as JSON:

{
  "contacts": [
    {
      "name": "string",
      "email": "string",
      "phone": "string or null",
      "company": "string or null",
      "title": "string or null"
    }
  ],
  "meeting_requested": "boolean",
  "urgency": "low | medium | high"
}

Return ONLY valid JSON with no additional text or explanation.

2

Request Valid, Parsable Output

LLMs sometimes add commentary before or after JSON. Be explicit about what you need.

❌ Bad Output:

Here's the JSON you requested:

{"name": "John"}

Hope this helps!

✓ Good Output:

{"name": "John"}

Add to prompt: "Output ONLY the JSON object with no additional text, explanation, or markdown formatting."

3

Handle Missing Data Gracefully

Specify how to handle missing or unavailable information in the output.

Include in your prompts:

  • • "If a field value is not found, use null (not empty string)"
  • • "If uncertain, use a confidence score field: "confidence": 0.0-1.0"
  • • "For arrays, return empty array [] if no items found, not null"

Tables & Lists

4

Define Table Structure

Specify column headers, data types, and the format (Markdown, CSV, HTML, etc.)

Compare these 5 project management tools and present as a Markdown table.

Columns (in this order):

  • Tool Name
  • Starting Price (format: "$XX/user/mo" or "Free")
  • Best For (max 5 words)
  • Key Feature
  • Rating (X.X/5.0)

Format as a Markdown table with proper alignment. Add a row separator between headers and data.

5

Specify List Format and Order

Numbered, bulleted, or nested? Sorted by what criteria? Be specific.

Examples:

"Return as a numbered list, sorted by priority (highest first)"

"Format as nested bullet points: main categories as bullets, sub-items indented with dashes"

"Create a checklist with [ ] for incomplete items"

Markdown Formatting

6

Specify Markdown Elements

Request specific Markdown features for properly formatted, readable output.

Write a technical guide and format as Markdown with these requirements:

  • Use ## for main sections, ### for subsections
  • Code blocks with language tags (```javascript, ```python, etc.)
  • Bold for important terms on first use
  • Inline code (`backticks`) for commands, file names, variables
  • Blockquotes (>) for warnings or important notes
  • Links in format [text](url) for external references

API-Friendly Responses

7

Design for Programmatic Consumption

When building AI-powered APIs, structure responses for easy parsing and error handling.

Analyze customer sentiment and return response in this format:

{
  "status": "success",
  "data": {
    "sentiment": "positive | neutral | negative",
    "confidence": 0.95,
    "score": -1.0 to 1.0,
    "key_themes": ["string", "string"],
    "suggested_action": "string"
  },
  "metadata": {
    "model": "claude-3-opus",
    "processing_time_ms": 234
  }
}
8

Include Error States

Define how the LLM should respond when it can't fulfill the request.

Add to prompts:

"If you cannot complete the task, return:

{
  "status": "error",
  "error_code": "INSUFFICIENT_DATA | AMBIGUOUS_REQUEST | etc",
  "message": "Human-readable error description",
  "suggestions": ["What user should provide"]
}

Do not make up data or guess when information is missing."

Output Formatting Best Practices

DO: Provide Output Examples

Show 1-2 examples of the exact format you want

DO: Validate Programmatically

Parse the output to verify it's valid JSON/XML/CSV before using it

DO: Use Delimiters

For complex prompts, use delimiters like ```json...``` to separate output from reasoning

DO: Specify Data Types

Be explicit: "string", "number", "boolean", "array", "ISO 8601 date", etc.

DON'T: Assume Default Formats

"Return as JSON" is too vague. LLMs may add markdown code blocks or explanations

DON'T: Ignore Edge Cases

Define how to handle empty results, nulls, very large outputs, and special characters

Build Robust AI Integrations

We can help you design API-ready prompts and workflows that produce consistent, parsable output

Schedule a Consultation Back to Home