JSON Prompting

Introduction

JSON prompting is when you ask the model to output JSON, often validated against a schema. It's the most production-critical formatting style because downstream code depends on it.

Beginner analogy: Like filling a database form — typed fields, no missing keys, predictable parse.

Understanding the topic

Core concepts to understand:

• Provide a JSON schema or example object.
• Use provider 'JSON mode' or 'structured outputs' for guarantees.
• Validate with Zod / Pydantic and retry on failure.
• Never wrap JSON in markdown when downstream will parse it.
• Stream JSON with care — partial JSON requires special parsing.

Syntax reference

Visual workflow / architecture:

Prompt:
"Respond with valid JSON only:
{
  \"name\": string,
  \"priority\": \"low\" | \"med\" | \"high\"
}
No prose, no markdown."

Real-world use

Every API-integrated LLM (Zapier, Make, custom backends) uses JSON prompts at the boundary between LLM and code.

Best practices

• Always validate; never trust the model's first JSON.
• Use provider structured-output APIs when available.
• Show an example object alongside the schema.

Common mistakes

• Asking for JSON but allowing 'commentary' — model adds prose.
• Forgetting to escape strings the model will output.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How do you guarantee valid JSON from an LLM?
• Q2. What's the difference between JSON mode and structured outputs?
• Q3. How do you handle a JSON parse failure in production?
• Q4. When NOT to use JSON output?