As AI-generated code and LLM-integrated APIs become foundational in software development, traditional JSON Schema shows its limits. This article introduces JSON Structure, a modern, strongly typed schema language designed for the era of AI. We explore how it improves data modeling, supports clean code generation (especially for TypeScript), and enables AI systems to reason more accurately about structured output.

TypeScript.TV

JSON Structure emerges as a modern alternative to JSON Schema, designed specifically for AI-driven development workflows. Unlike JSON Schema's validation-focused approach, JSON Structure offers strongly typed data definitions with precise numeric types (int32, uint64, decimal), native support for collections like sets and maps, and semantic extensions for physical units and currencies. The new IETF standard enables better code generation, cleaner inheritance patterns through $extends, and more accurate AI output validation, making it particularly valuable for LLM-integrated APIs and structured data generation.

Goodbye JSON Schema: Typing JSON with JSON Structure

<p>The article does not really provide any arguments why I should switch to JSON Structure. The provided examples are too shallow and niche (many domains just don’t need physical units). The benefits for inheritance are only hinted at and why it is better for LLMs is also not clear to me.</p>
<p>None of that means that JSON Structure is not good or better, just that the article leaves me with the XKCD “now there are 15 standards” feeling. Just to give one example: It’s true that JSON Schema has no support for UUIDs. There is a general suggestion to provide this with a “format” property. Why not build on that instead of creating something entirely new?</p>
