Output Formatting

Introduction

Controlling output format is the difference between an LLM that 'sometimes works' and one that ships. JSON, markdown, tables, fixed schemas — all are unlocked through the format section of your prompt.

Beginner analogy: Restaurants give chefs plating diagrams, not free reign. Format prompts plate the model's output the same way every time.

Understanding the topic

Core concepts to understand:

• Specify format explicitly (JSON, markdown, CSV, custom).
• Provide a schema or template if possible.
• Show ONE example of the exact format.
• Use 'response_format' parameter or JSON mode on supported models.
• Validate output programmatically (Zod, Pydantic) and retry on failure.

Syntax reference

Visual workflow / architecture:

Prompt:
"Return JSON exactly matching:
{
  \"summary\": string,
  \"tags\": string[],
  \"sentiment\": \"positive\" | \"negative\" | \"neutral\"
}"

Real-world use

Every LLM that integrates with a backend uses strict output formatting — invoice parsers, lead scorers, ticket triagers. Without it, downstream code breaks.

Best practices

• Validate output and retry once with the error appended to the prompt.
• Use providers' built-in JSON modes when available.
• Avoid asking for markdown when you'll parse it as JSON later.

Common mistakes

• Allowing 'or anything else you think is relevant' — model goes wild.
• Mixing markdown with JSON — parser breaks.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How do you guarantee JSON output from an LLM?
• Q2. What's a safe retry strategy on format failure?
• Q3. JSON mode vs prompt-based formatting — pros & cons?
• Q4. How do you handle markdown vs raw text?