Prompting and structured outputs

Prompting is not just writing instructions for a model. In a workflow, a prompt is also an interface between unstructured language and structured automation.

If the next node expects a field, list, label, or JSON object, design the prompt and output parser together.

Prompt anatomy

flowchart TD
  Role["Role or perspective"] --> Task["Task"]
  Task --> Context["Input context"]
  Context --> Rules["Rules and constraints"]
  Rules --> Output["Output format"]
  Output --> Examples["Optional examples"]

A good workflow prompt usually includes:

The task to perform.
The input fields the model should use.
What not to do.
The required output format.
A fallback when the answer is not available.

Weak vs strong prompts

Weak prompt	Stronger workflow prompt
“Summarize this.”	“Summarize `pageText` in 5 bullet points for a sales researcher. Do not include claims not present in the text.”
“Extract products.”	“Return JSON with `name`, `price`, `currency`, and `sourceUrl`. If a price is missing, set `price` to null.”
“Is this relevant?”	“Classify as `relevant` or `not_relevant`, then provide one sentence explaining the decision from the source text.”

Structured output pattern

Use structured output when another node must consume the AI result.

flowchart LR
  Text["Page text"] --> Prompt["Prompt with schema"]
  Prompt --> Model["Chat model"]
  Model --> Parser["Structured Output Parser"]
  Parser --> Next["Filter / integration / file"]

  style Parser fill:#e8f5e9,stroke:#2e7d32

Plain prose is fine for humans. JSON-like output is better for workflows.

Design the schema first

Before writing the prompt, decide what the next node needs.

{
  "companyName": "string",
  "pricingMentioned": "boolean",
  "price": "number or null",
  "currency": "string or null",
  "sourceQuote": "string"
}

Then write the prompt around that structure:

Extract pricing information from pageText. Return only the fields in the schema. If the page does not mention pricing, set pricingMentioned to false and price to null. Use a short source quote from the page when available.

Fallbacks are part of the output

Do not make the model guess silently. Give it a valid output for missing information.

Situation	Better fallback
Source does not mention a value	`null` plus a reason
Classification is uncertain	`unknown` with confidence
Retrieved context is insufficient	“Not enough source information”
Tool call fails	Structured error field

Test prompts with edge cases

Test with:

Empty input.
Very long input.
Missing fields.
Conflicting source text.
Pages with navigation text mixed into content.
Non-English or unusually formatted content.