LLM-to-JSON extraction prompt with schema, validation and retries

Prompt

Act as an LLM engineer specializing in structured data extraction for production systems.

Context:
- Input document type: {{document}}
- Fields I need to extract (name, type, required or not): {{fields}}
- Language and validation library: {{language}}
- A real input sample (anonymized):
```
{{sample}}
```

Deliverables:
1. A complete JSON Schema with types, formats (ISO 8601 dates, amounts as numbers with currency as a separate field), enums and required fields.
2. A system prompt for extraction with explicit rules: return only valid JSON, use null when a value does not appear (never infer it), do not normalize proper names, how to handle multiple candidate values and ambiguous text.
3. Two input/output examples (few-shot), one of them with missing fields.
4. Code in {{language}} that calls the model, enables structured output or JSON mode if the provider offers it, validates against the schema and retries at most twice, sending back the validation error.
5. An optional confidence or evidence field (a verbatim quote from the text) for each value, and when it is worth using.
6. Ten test cases: incomplete documents, different date formats, mixed languages, contradictory values and malicious instructions inside the text.
7. How to measure quality: per-field accuracy against a hand-labeled set.

Constraints:
- Do not invent API parameters: if an option depends on the provider, mark it [VERIFY in the official docs].
- The model must never fill a field without evidence in the text.
- If the fields are ambiguous or the sample is missing, ask me first.

Format: numbered sections with JSON and {{language}} code blocks.

How to use it

Usage example

Sample output

from enum import Enum
from pydantic import BaseModel, Field, ValidationError

class RenovationType(str, Enum):  # final values: [FILL IN]
    bathroom = "bathroom"
    kitchen = "kitchen"
    whole_home = "whole_home"
    other = "other"

class QuoteRequest(BaseModel):
    name: str
    phone: str | None = None
    city: str | None = None
    renovation_type: RenovationType | None = None
    max_budget_usd: float | None = Field(default=None, ge=0)

def extract(text: str, max_retries: int = 2) -> QuoteRequest:
    messages = [{"role": "user", "content": text}]
    for _ in range(max_retries + 1):
        # JSON mode / structured output depends on the provider: [VERIFY in the official docs]
        raw = call_model(SYSTEM_PROMPT, messages)
        try:
            return QuoteRequest.model_validate_json(raw)
        except ValidationError as error:
            messages += [
                {"role": "assistant", "content": raw},
                {"role": "user", "content": f"Invalid JSON: {error}. Return only the corrected JSON."},
            ]
    raise ValueError("The model did not return valid JSON")

{"name": "Megan", "phone": "555-0134", "city": "Denver",
 "renovation_type": "bathroom", "max_budget_usd": 9000}

Key facts

Category

Programming & development · AI agents

Output type

Code

Level

Advanced

Designed for

ChatGPT, Claude and Gemini

Verified in

Pending editorial verification

Variables

Document type, Fields to extract, Language and validation and Input sample

Language

English

Price

Free, no sign-up

Updated

Nov 7, 2026

Frequently asked questions