Schemas make LLM output trustworthy. Prompts do not.

Asking politely has a ceiling

A common approach is to write this in the prompt:

Reply only with JSON in the format below. Do not add anything else.

It works. About 99% of the time. The remaining 1% grows with your traffic.

Enforce it with structure

Making the model follow a format and refusing output that does not follow it are different things. The second one is code, not prose.

record ExtractedInvoice(String vendor, LocalDate issuedOn, long totalAmount) {}

ExtractedInvoice result = client.structured(prompt, ExtractedInvoice.class);

Now a format violation is a parse failure, and a parse failure takes a defined path: retry, or human review. That is a different kind of guarantee than one more "must" in the prompt.

A narrower schema shortens the prompt

Field names and types already carry intent. totalAmount: long says "digits only, no currency symbol" without a sentence saying so.