Structured outputs

Asking for JSON in the prompt and hoping is not a strategy. You get JSON most of the time, then a markdown code fence, then a helpful preamble, then a trailing comma, and your parser throws on input that was substantively correct.

Structured outputs make the shape a constraint rather than a request. You supply a schema, and the response conforms to it.

The shape

import { generateObject } from 'ai'
import { z } from 'zod'

const { object } = await generateObject({
  model: openai('gpt-5-mini'),
  schema: z.object({
    sentiment: z.enum(['positive', 'neutral', 'negative']),
    confidence: z.number().min(0).max(1),
    reasons: z.array(z.string()).max(3),
  }),
  prompt: review,
})

object.sentiment   // typed, parsed, and within the enum

The enum is doing real work: sentiment cannot come back as "quite positive" or "POSITIVE". You have made an entire category of downstream defensive code unnecessary.

Where it changes the most

• Anything you branch on. A classification you route with must be one of a fixed set, or the routing has an undefined case.
• Evaluation. A rubric score has to be aggregatable, which means parseable, which means schema-constrained.
• Chained steps. In prompt chaining, one step's output is the next step's input. Loose shapes compound.

What it does not fix

A schema constrains form, not truth. The model will happily return a perfectly valid object with a confidently wrong sentiment, and confidence: 0.95 is a generated number, not a calibrated probability. Schema validation passing tells you the response is readable, not that it is right. That is still the job of evaluation.