Generate Structured Output Schema

Generate Structured Output Schema for OpenAI, Gemini, Claude, and MCP from a JSON sample. Runs locally. Nothing is uploaded.

Free text is hard to parse. Constrain the model to return JSON by field before you argue about prompts. This page wraps a sample in each platform's shell.

There should be one core Schema: lock types, required, and additionalProperties, then pick OpenAI / Gemini / Claude / MCP.

Picking a platform only changes the local wrapper. The sample is not sent to that vendor.

Why strict mode fails easily

OpenAI strict usually requires objects to declare additionalProperties: false, and properties that appeared go into required. Check those two first, then try a real reply.

This is not the same as Function Calling

Structured Output constrains the model's final reply. Function Calling constrains tool input. Agents use both — don't force them into one Schema.

How to use it

Paste a JSON sample or type placeholders and pick a target platform.
Click Generate Schema and check that wrapper fields match the platform.
Add description, tighten enum, and confirm the extra-fields policy.
Validate a second response against the Schema. Don't stop at "it parses."
Copy it into the app. When you switch platforms, switch only the wrapper — don't fork two core structures.

When to use it

Move LLM output from Markdown to parseable JSON.
Add a response schema for extract, classify, or fill-in-a-form tasks.
Export the same core structure for OpenAI and Gemini side by side.
Produce the Schema first, then align frontend and backend with Zod / OpenAPI.

Keep in mind

Auto inference is optimistic. Don't treat it as the final production draft.
Numeric strings in the sample become string. Use real numbers for prices and quantities.
Empty arrays and mixed-type arrays widen items. Clean the sample before you generate.

FAQ

What should I watch in strict mode?

OpenAI strict usually requires objects to declare additionalProperties: false and a clear required. Check both after generate.

Can I use this as the production Schema?

No. Local inference is optimistic. Accept it with a real response.

How is this different from Function Calling?

Structured Output constrains the model's final reply. Function Calling constrains tool input.

Is my data uploaded?

No. Picking a platform only changes the local wrapper format.

Can Gemini and OpenAI share one Schema?

Core properties can be shared. Wrapper fields differ — don't drop json_schema onto Gemini as-is.

Should I write description?

Yes. The model understands field meaning from descriptions. type alone is not enough.

How do I constrain objects inside an array?

Keep element shapes consistent in the sample, then check items.properties and required.

How does this relate to the blog post?

The blog covers acceptance order and pitfalls. This page generates a first draft from a sample.