Skill עיבוד מסמכים עם AI ל-Claude

מתי להשתמש

"OCR", "Document extraction", "PDF AI", "Receipt scanning", "Invoice processing", "Vision API".

הוראות עבודה

1. Use Cases

High-ROI Examples

• Invoice processing — extract vendor/amount/date.
• Receipt management — auto-categorize expenses.
• ID verification — KYC.
• Contract review — extract clauses.
• Resume parsing — extract structured data.
• Form digitization — paper → DB.
• Medical records — structured patient data.

2. Tools — 2026

3. Approach Decision

Simple OCR (text → text)

• Tesseract (free, open source).
• Google Vision OCR (cheap, fast).

Structured extraction (typed fields)

• Claude Vision if smart reasoning needed.
• AWS Textract for forms/tables.
• Mindee/Klippa for specialized receipts.

Classification + Extraction

• Claude Vision = best.
• Pre-trained models (Mindee, Document AI) for specific types.

4. Sample Workflow — Invoice Processing

1. Trigger: Email with PDF attachment
2. Extract attachment
3. Convert PDF → Image (if needed)
4. Claude Vision:
   System: "You are an invoice processor."
   User: [Image] + "Extract: vendor, invoice_number, date, amount, line_items, due_date as JSON."
5. Validate output (schema check)
6. Match to PO if exists
7. Decision routing (approval workflow)
8. Add to QuickBooks
9. Notify accounting team

5. Hebrew Document Processing

OCR Tools that Support Hebrew

• Google Vision API ✅ best.
• AWS Textract ✅ basic.
• Tesseract ✅ with Hebrew model.
• Claude Vision ✅ excellent.

Israeli Use Cases

• Israeli tax forms (101, 102, 161).
• Israeli receipts (חשבוניות).
• Hebrew contracts.
• בנקאות docs.

6. Prompt Design for Document AI

You are a [doc type] processor.

Extract the following fields from the document:
- field1 (type, e.g., string)
- field2 (number)
- field3 (date YYYY-MM-DD)
- line_items (array of objects)

Rules:
- Use exact text from document for strings.
- Numbers without currency symbols.
- Dates in ISO format.
- If field missing, use null.
- If unclear, lower confidence in confidence field.

Respond in JSON only:
{
  "field1": "value",
  "field2": 123,
  "confidence": 0.95
}

7. Validation

Always

• Schema check — required fields present?
• Type check — number is number, date is valid?
• Range check — amount > 0, date plausible?
• Cross-validate — sum of line items = total?

Action on Low Confidence

• Confidence < 0.8 → human review.
• Confidence >= 0.8 → auto-process.

8. Handling Multi-Page Documents

• Split PDF to pages.
• Process each separately.
• Aggregate results.

Or

• Multi-page capable models (Document AI, Claude Vision).

9. Common Pitfalls

❌ Trust AI output blindly — always validate. ❌ No confidence threshold — bad data slips in. ❌ Hebrew handled like English — character flips, RTL issues. ❌ PDFs as one giant image — slower + worse quality. ❌ No human-in-the-loop for high-value docs.

10. Privacy + Compliance

• PII extraction — GDPR considerations.
• Israeli Privacy Law: Process locally if possible.
• Self-hosted options: Mindee on-prem, Tesseract.
• Data retention: Delete after processing.

11. Cost Optimization

Strategies

• OCR first (cheap), then LLM only on extracted text — cheaper.
• Cache repeat processing.
• Batch when possible.
• Resize images — smaller = cheaper for vision.
• Preprocess — better OCR = less LLM cleanup.

12. Sample Cost Calculation

1,000 invoices/month

• Approach 1 (Claude Vision direct): 1,000 × ~$0.01 = $10/m.
• Approach 2 (Mindee): 1,000 × $0.30 = $300/m.
• Approach 3 (Textract + Claude text): $1.50 + $5 = $6.50/m.

Approach 1 wins for small volume + smart extraction needed.

13. אסיים בהמלצה.

קלט נדרש

פלט צפוי

דגלים אדומים

• 🚨 No validation — bad data downstream.
• 🚨 Auto-approve high $$$ without review.
• 🚨 PII to public APIs — privacy violation.
• ⚠️ Single-page approach for multi-page — context loss.

הערות חשובות

• Claude Vision = best general-purpose.
• Specialized tools (Mindee) = better for high-volume specific docs.
• Tesseract for free but worse than Vision APIs.

פרומפט לדוגמה

Israeli company, 500 invoices/month in Hebrew. Tool?

Resume parsing workflow ב-Make.

Receipt extraction for expense reports. ROI?

© 2026 Automation Expert Pro | גרסה 1.0.0