Force JSON Output from AI

Overview

Asking for JSON politely fails in three ways: the model wraps the object in prose, fences it in markdown, or invents fields you never asked for. Forcing JSON reliably takes a contract, not a request: a schema with explicit types and requirements, a realistic example response the model can imitate, validation rules for nulls and types, and a strict rule block that bans everything except the object itself. This resource loads a complete strict-mode contract so you can see all four parts working together — then adapt the schema to your own data.

Workflow

1. 1
   Open the loaded contract

   The customer-record schema loads in Strict mode. Generate and read the four sections — each handles a different failure mode.

2. 2
   Swap in your fields

   Replace the schema fields with your own; the example response regenerates with realistic values automatically.

3. 3
   Test at Standard first

   If your model behaves, Standard's four rules cost fewer tokens. Escalate to Strict only when prose leaks through.

Why This Works

• The example response does more than the rules — models imitate examples more faithfully than they follow instructions
• Explicit null discipline kills the two worst failure modes: omitted keys and invented values
• The strict block addresses each leak (prose, fences, commentary, extra objects) by name instead of saying 'only JSON' and hoping

Use cases

• Making a chat model's output safe to JSON.parse without cleanup code
• Hardening an API integration that breaks on prose-wrapped responses
• Replacing 'return valid JSON' one-liners with a contract that actually holds