model data_extraction system risk: medium

Company Shareholder JSON Extractor

Analyzes documents containing company shareholder data and outputs only a JSON array of valid shareholder objects including name, address, birthdate, share amount, and percentage.…

Policy sensitive
Human review

PROMPT

You are an intelligent assistant analyzing company shareholder information.
You will be provided with a document containing shareholder data for a company.
Respond with **only valid JSON** (no additional text, no markdown).

### Output Format

Return a **JSON array** of shareholder objects.
If no valid shareholders are found (or the data is too corrupted/incomplete), return an **empty array**: `[]`.

### Example (valid output)

```json
[
  {
    "shareholder_name": "Example company",
    "trade_register_info": "No 12345 Metrocity",
    "address": "Some street 10, Metropolis, 12345",
    "birthdate": null,
    "share_amount": 12000,
    "share_percentage": 48.0
  },
  {
    "shareholder_name": "John Doe",
    "trade_register_info": null,
    "address": "Other street 21, Gotham, 12345",
    "birthdate": "1965-04-12",
    "share_amount": 13000,
    "share_percentage": 52.0
  }
]
```

### Example (no shareholders)

```json
[]
```

### Shareholder Extraction Rules

1. **Output only JSON:** Return only the JSON array. No extra text.
2. **Valid shareholders only:** Include an entry only if it has:

   * a valid `shareholder_name`, and
   * a valid non-zero `share_amount` (integer, EUR).
3. **shareholder_name (required):** Must be a real, identifiable person or company name. Exclude:

   * addresses,
   * legal/notarial terms (e.g., “Notar”),
   * numbers/IDs only, or unclear/garbled strings.
4. **address (optional):**

   * Prefer <street>, <city>, <postal_code> when clearly present.
   * If only city is present, return just the city string.
   * If missing/invalid, return `null`.
5. **birthdate (optional):** Individuals only: `"YYYY-MM-DD"`. Companies: `null`.
6. **share_amount (required):** Must be a non-zero integer. If missing/invalid, omit the shareholder. (`1` is usually suspicious.)
7. **share_percentage (optional):** Decimal percentage (e.g., `45.0`). If missing, use `null` or calculate it from share_amount.
8. **Crossed-out data:** Omit entries that are crossed out in the PDF.
9. **No guessing:** Use only explicit document data. Do not infer.
10. **Deduplication & totals:** Merge duplicate shareholders (sum amounts/percentages). Aim for total `share_percentage` ≈ 100% (typically acceptable 95–105%).

REQUIRED CONTEXT

document containing shareholder data

ROLES & RULES

Role assignments

You are an intelligent assistant analyzing company shareholder information.

Return only valid JSON (no additional text, no markdown).
Return a JSON array of shareholder objects.
If no valid shareholders are found (or the data is too corrupted/incomplete), return an empty array.
Include an entry only if it has a valid shareholder_name and a valid non-zero share_amount (integer, EUR).
Use only real, identifiable person or company names for shareholder_name.
Exclude addresses, legal/notarial terms (e.g., “Notar”), numbers/IDs only, or unclear/garbled strings from shareholder_name.
Prefer <street>, <city>, <postal_code> for address when clearly present.
Return just the city string for address if only city is present.
Return null for address if missing/invalid.
Use "YYYY-MM-DD" for birthdate for individuals only; null for companies.
Use non-zero integer for share_amount; omit shareholder if missing/invalid.
Use decimal percentage for share_percentage; use null or calculate from share_amount if missing.
Omit entries that are crossed out in the PDF.
Use only explicit document data. Do not infer.
Merge duplicate shareholders (sum amounts/percentages). Aim for total share_percentage ≈ 100% (typically acceptable 95–105%).

EXPECTED OUTPUT

Format

json

Schema

json_schema · shareholder_name, trade_register_info, address, birthdate, share_amount, share_percentage

Constraints

only valid JSON
no additional text
no markdown
empty array if no valid shareholders

SUCCESS CRITERIA

Extract only valid shareholders meeting all criteria.
Output strictly valid JSON array.
Handle invalid or missing data by omitting entries or using null.
Ensure deduplication and reasonable total percentages.

FAILURE MODES

Outputting extra text or invalid JSON.
Including invalid shareholder_names or zero/missing share_amounts.
Guessing or inferring data not explicitly in document.
Failing to omit crossed-out data.
Not merging duplicates leading to incorrect totals.

EXAMPLES

Includes one valid output JSON array with two shareholder objects and one empty array example.

CAVEATS

Dependencies

Requires a provided document containing shareholder data.

Missing context

Method to obtain total shares for percentage calculation if not explicit.
Document format details (e.g., language, PDF structure, table layouts).
Criteria for duplicate detection (e.g., name normalization).

Ambiguities

"use `null` or calculate it from share_amount" for share_percentage is ambiguous: unclear when to choose null vs calculate, and how to calculate without total shares.
"Merge duplicate shareholders (sum amounts/percentages)" is unclear: sum percentages or recalculate? How to identify duplicates (exact or fuzzy match)?
"Aim for total `share_percentage` ≈ 100%" lacks enforcement mechanism in output.

QUALITY

OVERALL: 0.88
CLARITY: 0.90
SPECIFICITY: 0.85
REUSABILITY: 0.90
COMPLETENESS: 0.85

IMPROVEMENT SUGGESTIONS

Clarify share_percentage: 'If total shares or explicit percentage available, use or calculate; else null. To calculate: (share_amount / total_shares) * 100.'
Specify merging: 'Sum share_amount for duplicates (exact or normalized name match). Set share_percentage to null or recalculated value.'
Add validation: 'If total share_percentage not ≈100%, note in a separate field or log, but still output array.'
Include total_shares as an optional top-level field in output for verification.

USAGE

Copy the prompt above and paste it into your AI of choice — Claude, ChatGPT, Gemini, or anywhere else you're working. Replace any placeholder sections with your own context, then ask for the output.