Skill Claude API ב-Automations ל-Claude

מתי להשתמש

"Claude API", "Anthropic", "Claude in workflows", "Tool use", "Prompt caching".

הוראות עבודה

1. Why Claude vs GPT-4

Claude Strengths

• Long context (200K tokens) — full docs, conversations.
• Coding excellent.
• Following instructions precise.
• Hebrew strong.
• Lower hallucinations typically.

GPT-4 Strengths

• Image generation (DALL-E).
• Voice (Whisper, TTS).
• More integrations out of box.

2. Models — When to Use

Claude Opus 4

• Most powerful.
• Complex analysis, reasoning.
• $15 / $75 per 1M tokens (in/out).
• Use: Critical tasks, low volume.

Claude Sonnet 4

• Balanced.
• Most use cases.
• $3 / $15 per 1M tokens.
• Use: Default for production workflows.

Claude Haiku 4

• Fast + Cheap.
• Simple tasks.
• $0.25 / $1.25 per 1M tokens.
• Use: Classification, extraction, high volume.

3. Setup in Make.com

1. Module: Anthropic Claude — Make Completion
2. Connection: Add API key (from console.anthropic.com)
3. Configure:
   Model: claude-haiku-4-5 (or sonnet/opus)
   System prompt: "You are a..."
   User message: {{from previous step}}
   Max tokens: 1024
   Temperature: 0 (deterministic) to 1 (creative)

4. Tool Use — Function Calling

Concept

• Claude calls your functions/APIs based on user request.
• Like an agent.

Example

System: "You can use these tools:
- get_weather(city)
- send_email(to, subject, body)

User: 'What's the weather in Tel Aviv? Email Dana the report.'

Claude: I'll call get_weather('Tel Aviv') first.
[Returns: 22°C, sunny]

Claude: Now call send_email('dana@...', 'Tel Aviv Weather', 'Currently 22°C and sunny')."

Use in Workflows

• Automate multi-step tasks.
• Claude decides what to do based on input.

5. Prompt Caching — חיסכון 90%!

What

• Cache common prompt prefixes.
• 90% cheaper, faster.

When to Use

• Same system prompt across many calls.
• RAG context reused.
• Examples (few-shot) repeated.

How

# Mark cached portion
{
  "type": "text",
  "text": "[Long system prompt with examples]",
  "cache_control": {"type": "ephemeral"}
}

Cost

• Cache write: 1.25x normal.
• Cache read: 0.1x normal (10% cost).
• Cache lifetime: 5 minutes (refreshed on use).

6. JSON Mode

Why

• Reliable parsing in workflows.
• No "Sure, here's the JSON: {...}" preamble.

How

System: "Always respond in valid JSON."

User: "Extract name and email from: 'Contact Dana at dana@acme.co'"

Claude:
{
  "name": "Dana",
  "email": "dana@acme.co"
}

7. Vision (Image Analysis)

Use Cases

• Receipt OCR.
• Product image classification.
• Document analysis.
• Screenshot analysis.

How

User: [Image of receipt]
System: "Extract: vendor, date, amount, items."

Claude: {
  "vendor": "Café Aroma",
  "date": "2026-05-07",
  "amount": "₪52.00",
  "items": ["Cappuccino", "Croissant"]
}

8. Batch API — חיסכון 50%

What

• Send 1,000s of requests in 1 call.
• Process within 24 hours.
• 50% discount.

When

• Async OK (not real-time).
• Bulk processing.

Use Cases

• Re-classify 10K old tickets.
• Generate descriptions for 1K products.
• Translate 5K reviews.

9. Sample Workflow — Lead Enrichment

1. Trigger: New lead
2. HTTP: Fetch company website HTML
3. Claude (Sonnet): Extract:
   - Industry
   - Company size estimate
   - Recent news
   - Tech stack mentioned
   - Likely buyer persona
   System: "You are a B2B sales researcher. Extract structured data."
   Max tokens: 500
4. Update CRM with extracted fields
5. Slack notification with summary

10. Cost Optimization

Strategies

• Use cheapest model that works.
• Cache common prompts.
• Batch when possible.
• Limit max_tokens.
• Avoid retries on success.

Monitor

• Console.anthropic.com → Usage.
• Set spending limits.
• Alert on anomalies.

11. Israel Specifics

• Hebrew prompts: excellent quality.
• Hebrew output: formats reliably.
• Privacy: Anthropic doesn't train on API data.
• Cost in USD: ₪3.7-3.8 = $1.

12. אסיים בהמלצה.

קלט נדרש

פלט צפוי

דגלים אדומים

• 🚨 Always Opus — overkill, expensive.
• 🚨 No max_tokens — runaway cost.
• 🚨 No JSON mode — parsing fails.
• ⚠️ No prompt caching — leaving money on table.

הערות חשובות

• Anthropic console: Test prompts before integrating.
• API docs: docs.anthropic.com — excellent.
• Claude Sonnet 4.6 = workhorse model.

פרומפט לדוגמה

Lead enrichment workflow ב-Make. Use Claude.

Classify 10K tickets — Batch or real-time?

Vision: extract data from receipts. Build it.

© 2026 Automation Expert Pro | גרסה 1.0.0