> ## Documentation Index
> Fetch the complete documentation index at: https://docs.famulor.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Prompt Writing Basics

> Learn how to write effective system prompts and initial messages for your AI assistants

famulor.de uses a **System Prompt** to guide how your AI assistant interacts with callers. Crafting an effective prompt is key to smoother, more accurate phone conversations.

## 1. System Prompt vs. Initial Message

### What is a System Prompt?

The **System Prompt** contains overall instructions and context that guide your AI assistant's behavior throughout the entire conversation. It defines:

* The assistant's role and personality
* How to handle different scenarios
* What information to collect
* When to transfer or end calls

**Example System Prompt:**

> "You are the Support Bot for TechMasters Inc. Greet politely, ask for a name, handle basic troubleshooting, and transfer the call if the issue is complex."

### What is an Initial Message?

The **Initial Message** is the very first thing the AI says when answering a call. It's especially important for inbound calls as it sets the tone and expectations.

**Example Initial Message:**

> "Thank you for calling TechMasters! How can I assist you today?"

<Info>
  The Initial Message is separate from the System Prompt. The System Prompt guides behavior, while the Initial Message is the actual greeting callers hear.
</Info>

## 2. How to Write Clear and Effective Prompts

### Be Specific About Purpose and Tone

Clearly outline the main purpose, desired tone, and step-by-step process you want the AI to follow. Avoid vague instructions.

**Good Example:**

> "You are a friendly customer service representative for XYZ Company. When a caller mentions a billing question, ask for their invoice number (format: INV-XXXXX), then look up their account status."

**Bad Example:**

> "Help customers with questions."

### Use Concrete Examples

Show sample dialogues and specific scenarios to guide the AI's responses.

**Example:**

> "If the user says they have a billing question, ask them for their invoice number. If they say 'I don't have it,' politely ask for their email address or phone number on file instead."

### Structure Your Prompt with Clear Sections

Organize your prompt with headings or numbered steps to make it easier for the AI to follow:

1. **Role Definition**: Who is the assistant?
2. **Tone Guidelines**: How should they speak?
3. **Common Scenarios**: What situations will they handle?
4. **Data Collection**: What information do you need?
5. **Escalation Rules**: When to transfer or end calls?

## 3. How to Handle Unknown or Off-Topic Questions

When callers ask about topics not covered in your prompt, the AI needs clear instructions on how to respond.

### Option 1: Default Response with Transfer

Provide a polite default response and offer to transfer:

**Prompt Example:**

> "If a caller asks about something you don't know, say: 'I'm not sure about that specific question. Let me transfer you to a specialist who can help.' Then use the transfer tool."

### Option 2: Gather More Information

Ask follow-up questions to better understand the request:

**Prompt Example:**

> "If a caller asks about something unclear, ask: 'Could you provide a bit more detail so I can help you better?' Then try to route them based on their response."

### Option 3: Reference Knowledge Base

If you have a knowledge base configured, instruct the AI to search it:

**Prompt Example:**

> "For questions you're unsure about, first check the knowledge base. If no relevant information is found, politely transfer the caller to a human agent."

<Tip>
  Test your prompt with edge cases by reviewing call transcripts. Adjust your instructions based on how the AI handles unexpected questions.
</Tip>

## 4. How Long Should Your Prompt Be?

### Optimal Prompt Length

* **Short prompts (50-200 words)**: Work best for simple use cases with clear, direct instructions
* **Medium prompts (200-500 words)**: Good for handling multiple scenarios with structured guidelines
* **Long prompts (500+ words)**: May slow down responses and increase "hallucination" risk

### When to Use a Knowledge Base Instead

If you have extensive information to share (product catalogs, detailed FAQs, company policies), use a **Knowledge Base** instead of putting everything in the system prompt:

* Knowledge bases are searchable and context-aware
* They don't slow down the AI's response time
* They can be updated without changing the prompt
* They handle large amounts of information more efficiently

Learn more: [Knowledge Bases Guide](/en/conversation-design/knowledge-bases)

### Best Practices for Prompt Length

1. **Start Short**: Begin with a concise prompt and add details only as needed
2. **Use Structure**: Break long prompts into clear sections with headings
3. **Test Performance**: Monitor call latency - if responses slow down, consider moving content to a knowledge base
4. **Review Regularly**: Remove unnecessary instructions that don't improve outcomes

<Warning>
  Very long prompts (1000+ words) can cause slower response times and inconsistent behavior. Consider using knowledge bases for detailed information.
</Warning>

***

## 5. Testing and Improving Your Prompts

### Review Call Transcripts Regularly

Periodically review call transcripts to identify:

* Questions the AI couldn't answer
* Responses that don't match your brand voice
* Scenarios where the AI got confused
* Opportunities to add or remove instructions

### Iterative Improvement Process

1. **Deploy** your initial prompt
2. **Monitor** call transcripts and analytics
3. **Identify** patterns in failures or misunderstandings
4. **Refine** your prompt with specific examples
5. **Test** changes with a few calls before full rollout

### Common Prompt Issues and Solutions

| Issue                            | Solution                                              |
| -------------------------------- | ----------------------------------------------------- |
| AI gives generic responses       | Add specific examples and scenarios                   |
| AI doesn't collect required data | Explicitly list what information to gather            |
| AI transfers too early/late      | Define clear escalation criteria                      |
| AI sounds robotic                | Adjust tone instructions and add personality          |
| AI hallucinates information      | Add instructions to say "I don't know" when uncertain |

***

**Next Steps**

* Learn about [Knowledge Bases](/en/conversation-design/knowledge-bases) for handling extensive information
* Explore [Interruption Handling](/en/conversation-design/interruptions) for natural conversations
* Check [System Prompt Best Practices](/en/ai-assistants/system-prompt) for advanced techniques