Lesson 2 · 7 min

Structured Output

JSON mode, schema-guided generation, and post-validation.

For machine-readable output, give Claude an example *and* a schema. JSON-mode + Zod validation in code is the safest combo: model produces JSON, code parses and validates, retry on parse error with the validator's message.

Production scenario

Real-world example: Resume parser at an ATS

An applicant tracking system parses uploaded resumes into structured records. The recipe that finally stuck in production:

const schema = z.object({
  name: z.string(),
  email: z.string().email(),
  experience: z.array(z.object({
    company: z.string(),
    title: z.string(),
    start: z.string(),
    end: z.string().nullable(),
  })),
  skills: z.array(z.string()),
});

const parsed = await callWithRetry({
  system: SYSTEM,
  user: `Parse this resume into JSON matching the schema.\n${resumeText}`,
  schema,
  maxRetries: 1,
});

The retry appends the Zod error to the prompt: "Last attempt failed because: end date for 2nd job was not a valid string." The model fixes it on attempt two ~95% of the time.

Why this matters: JSON-mode reduces malformed output; schema validation catches what slips through; one retry on the validator error closes the gap.

Knowledge points in this lesson

Provide schema and an example
JSON mode reduces but doesn't eliminate parse errors
Validate with Zod after parse
Retry once with validator error appended
Use tool-call output for hardest cases
XML tags help when output has sections

Quick check

Prompt EngineeringSelect one

How long does Anthropic's prompt cache persist by default?