Skill עיצוב System Prompts ל-Claude

מתי להשתמש

"System prompt", "Persona design", "Custom GPT instructions", "Project instructions", "Behavior rules".

הוראות עבודה

1. What is a System Prompt

Persistent instructions that apply to every message in conversation. Set once, apply always.

Where Used

• Custom GPTs (instructions field).
• Claude Projects (custom instructions).
• API calls (system parameter).
• Plugins (CLAUDE.md).
• Skills (frontmatter + content).

2. System Prompt Structure

[1. Role Definition]
You are an expert at...

[2. Goals / Purpose]
Your primary goal is to...

[3. Behavioral Rules]
Always:
- Do X
- Be Y

Never:
- Do Z
- Reveal A

[4. Output Format]
Always respond in:
- Hebrew (unless asked otherwise)
- Markdown formatting
- Max 500 words

[5. Examples (optional)]
Example interaction:
User: ...
Assistant: ...

[6. Edge Cases]
If user asks about [X]:
- Respond: ...

[7. Tone & Style]
- Professional but warm
- Direct, no fluff
- Use specific numbers when possible

3. Length Recommendations

Diminishing returns after ~3,000 words for general use.

4. Role Definition

Weak

"You are a helpful assistant."

Strong

"You are a Senior Marketing Operations Director with 15 years experience scaling B2B SaaS from $1M to $50M ARR. You've worked at HubSpot, Drift, and 3 unicorns. You speak in concrete examples and actionable frameworks."

5. Behavioral Rules

Format

• ✅ "Always cite sources when stating facts"
• ✅ "Refuse to provide medical/legal advice — direct to professional"
• ✅ "If you're unsure, say so explicitly"

Don't be vague

• ❌ "Be helpful"
• ❌ "Be accurate"

6. Guardrails

Safety

• Refuse harmful requests.
• Refuse to bypass safety.
• Don't generate hate speech.

Brand

• Don't disparage competitors.
• Stay in scope.
• Don't make commitments.

Privacy

• Don't ask for PII.
• Don't store user data.

7. Output Formatting

Be Specific

• "Always respond in JSON: { 'category': '', 'confidence': 0-1 }"
• "Use markdown headings for sections > 100 words"
• "End with: 'Want me to dive deeper on any point?'"

Language

• "Respond in Hebrew unless asked otherwise"
• "Match user's language"

8. Examples in System Prompt

Example exchange:

User: How do I improve CAC?
You:
Great question. CAC has 3 levers:

1. **Lower acquisition cost**:
   - Negotiate ad rates
   - Improve organic
   
2. **Increase conversion**:
   - Landing page optimization
   - Better targeting

3. **Faster sales cycle**:
   - Sales enablement

Want me to dive deeper on any of these?

Examples set tone better than instructions alone.

9. Iterative Refinement

Process

1. Write v1.
2. Test with 10 real prompts.
3. Note where output is off.
4. Add rule / example to fix.
5. Test again.
6. Repeat until stable.

10. Common Mistakes

❌ Too short — "Be a marketing expert" → vague. ❌ Too long — 10K words → AI ignores middle. ❌ Conflicting rules — "Be brief" + "Cover all aspects". ❌ No examples — abstract instructions. ❌ No edge cases — ambiguous behavior.

11. Maintenance

• Version control system prompts.
• Document changes.
• A/B test major changes.
• Update as use case evolves.

12. אסיים בהמלצה.

פרומפט לדוגמה

Build system prompt for B2B SaaS sales coach.

System prompt too long (5K words). Optimize.

Test if my system prompt is working — how?

© 2026 AI Expert Pro | גרסה 1.0.0