JSON extraction prompt for structured data

Paste this into a chat window and replace each {{variable}}, or open it in the prompt builder to fill them in a form. Sections are marked with XML tags, which every major model reads well.

You extract structured data from text accurately. You never guess a value that the text doesn’t state.

<context>
<document>
{{document}}
</document>
</context>

<constraints>
- Use only information stated in the document. If a field isn’t there, use null. Don’t infer it from context or general knowledge.
- Copy names, IDs and amounts exactly as written. Normalise only where the shape says so: dates to YYYY-MM-DD, amounts to numbers without currency symbols.
- If a value is ambiguous (two different dates could be the due date), use null and explain in the "notes" field.
- Include every key in the shape, even when its value is null.
</constraints>

<output_format>
Only the JSON object. No code fences, no text before or after it.
</output_format>

<task>
Extract the fields below from the document and return them as JSON matching this shape:

{{json_shape}}
</task>

You extract structured data from text accurately. You never guess a value that the text doesn’t state.

<constraints>
- Use only information stated in the document. If a field isn’t there, use null. Don’t infer it from context or general knowledge.
- Copy names, IDs and amounts exactly as written. Normalise only where the shape says so: dates to YYYY-MM-DD, amounts to numbers without currency symbols.
- If a value is ambiguous (two different dates could be the due date), use null and explain in the "notes" field.
- Include every key in the shape, even when its value is null.
</constraints>

<output_format>
Only the JSON object. No code fences, no text before or after it.
</output_format>

<context>
<document>
{{document}}
</document>
</context>

<task>
Extract the fields below from the document and return them as JSON matching this shape:

{{json_shape}}
</task>

{{document}}

The source text: an email, an invoice, a PDF converted to text, a web page.

{{json_shape}}

The keys and types you want, as an example object or a JSON Schema. Include a "notes" string field for ambiguities.

{{document}}

From: accounts@northwind-supplies.example
Subject: Invoice NW-20931

Hi, please find invoice NW-20931 for October. Total due: £1,284.50 (incl. VAT).
Payment terms 30 days from the invoice date of 3 October 2026. PO ref: 77-1042.

{{json_shape}}

{
  "supplier_email": "string",
  "invoice_number": "string",
  "total": "number",
  "currency": "ISO 4217 code",
  "invoice_date": "YYYY-MM-DD",
  "due_date": "YYYY-MM-DD or null",
  "po_number": "string or null",
  "notes": "string or null"
}

What to look for. The due date isn’t written, only implied (30 days from 3 October). Under the rules it should be null with a note explaining the terms, rather than a computed 2 November that your accounting system might treat as stated fact.

You extract structured data from text accurately. You never guess a value that the text doesn’t state.

<context>
<document>
From: accounts@northwind-supplies.example
Subject: Invoice NW-20931

Hi, please find invoice NW-20931 for October. Total due: £1,284.50 (incl. VAT).
Payment terms 30 days from the invoice date of 3 October 2026. PO ref: 77-1042.
</document>
</context>

<constraints>
- Use only information stated in the document. If a field isn’t there, use null. Don’t infer it from context or general knowledge.
- Copy names, IDs and amounts exactly as written. Normalise only where the shape says so: dates to YYYY-MM-DD, amounts to numbers without currency symbols.
- If a value is ambiguous (two different dates could be the due date), use null and explain in the "notes" field.
- Include every key in the shape, even when its value is null.
</constraints>

<output_format>
Only the JSON object. No code fences, no text before or after it.
</output_format>

<task>
Extract the fields below from the document and return them as JSON matching this shape:

{
  "supplier_email": "string",
  "invoice_number": "string",
  "total": "number",
  "currency": "ISO 4217 code",
  "invoice_date": "YYYY-MM-DD",
  "due_date": "YYYY-MM-DD or null",
  "po_number": "string or null",
  "notes": "string or null"
}
</task>

• OpenAI GPT

  OpenAI’s Structured Outputs require every field in required; make a field optional with a union such as ["string", "null"], which fits the null-instead-of-guess rule exactly.

• Claude

  Anthropic’s structured outputs (output_config.format) constrain Claude’s reply to your schema. Without it, the format line is what replaces the old prefill trick, which Claude 4.6 models and later reject with a 400 error.

• Gemini

  Gemini takes a schema with response_format for structured output. Google warns that very large or deeply nested schemas may be rejected, so keep the shape flat where you can.

• Classify textSort text into fixed labels with definitions, examples and an “unclear” option, returning JSON.
• Clean and normalise dataNormalise messy records to your rules, flag the rows it can’t fix, and never drop or invent data.
• SQL from a schemaTurn a question in plain English into one query for your schema and SQL dialect, with assumptions stated.

Building this into an app that sends the same instructions with every request? Start from a system prompt instead; each comes with design notes, failure modes and a monthly cost. In the system prompt library: JSON-only extraction service.

Edited the template? Compare your version with this one in the prompt diff, and count it for your model with the token counter. In the glossary: structured outputs, hallucination. Browse all 29 prompt templates.