Structured outputs and multi-step workflows

Why structure the output

Structured outputs — tables, lists with fixed fields, JSON — are easier to compare, check and reuse. They let you paste results into a spreadsheet, feed them into another prompt or process them with a script.

Asking for structured output

Describe the structure exactly and show an example:

Extract every action item from the meeting notes below.
Return JSON only, as a list of objects with exactly these fields:
- "task": string, starting with a verb
- "owner": string, a role from the notes, or null if no owner is named
- "due": date as YYYY-MM-DD, or null if no date is given
Example: [{"task": "Book the venue", "owner": "Events lead", "due": "2025-03-14"}]
Notes:
<notes>
(paste notes)
</notes>

Key points: name every field, give types, say what to do when information is missing (null), and say "JSON only" if another tool will read it. Some AI platforms offer a "structured output" or "JSON schema" setting that enforces a format; use it when available, but still validate the result.

Validate before you trust

Even well-formed output can be wrong. Check:

• Shape — valid JSON? Every field present? Correct types?
• Content — does every value come from the source? Any invented owners or dates?
• Coverage — were any items missed? (Compare counts against the source.)

Multi-step workflows

A workflow connects several prompts (and sometimes tools) to complete a larger job. Design it on paper first:

1. Steps — what each prompt does, with its own instructions and output format.
2. Handoffs — what passes from one step to the next (only the approved output, not the whole conversation).
3. Checkpoints — where a human reviews before continuing, especially before anything is sent, published or decided.
4. Failure handling — what happens if a step's output fails validation (retry once, then escalate to a person).

Example: weekly customer-feedback digest

1. Classify each comment (few-shot classifier) → table.
2. Validate: every comment has exactly one label → if not, flag for review.
3. Summarise top themes per label with quoted evidence.
4. Human checkpoint — check themes against comments.
5. Draft a 150-word digest for the team using the approved themes.

Keep humans in the loop where it matters

Automating a workflow does not remove responsibility. Put review steps before outputs reach customers, affect people's records or money, or are published. Note in your workflow who reviews and what they check.