Prompt Guide for SipPulse AI Conversational Agents

Creating a prompt for a SipPulse AI Conversational Agent is fundamentally different from writing a prompt for a traditional LLM. Instead of asking for a single output, you're defining the personality, rules, and capabilities of an entity that will maintain continuous and contextual dialogue.

The "prompt" of a SipPulse AI Agent is its System Instruction (System Message). This is its constitution, the central directive it consults before each response to ensure it stays on track, uses tools correctly, and behaves as expected throughout the entire conversation.

The Right Mindset: From "Task" to "Dialogue"

The main mindset shift is moving from "Give me X" to "Be Y".

Characteristic	Traditional LLM Prompt	SipPulse AI Agent Instruction
Purpose	Get an answer to a single question.	Define behavior for an entire conversation.
Focus	On the quality of immediate output.	On persona consistency, dialogue flow, and tool execution.
State	Stateless (each prompt is new).	Persistent (instruction applies to each conversation turn).
Result	A block of text, code, or data.	A coherent and useful dialogue that may include actions (API calls, etc.).

Recommended Structure for Agent Instructions

A well-structured instruction is key to a reliable and effective agent. Dividing your instructions into clear sections helps the LLM better understand its role and responsibilities.

1. Persona and Role

This is the foundation of your agent's identity. Define who they are, what their tone is, and how they should present themselves.

• Example: "You are the 'SipPulse Assistant', a friendly, patient, and extremely helpful customer support specialist. Your tone should always be professional but empathetic. Never use slang or overly casual language."

2. Main Goal and Domain (Goal & Domain)

Clearly define the agent's purpose and the boundaries of its knowledge. What should it do and, equally important, what should it not do?

• Example: "Your main goal is to help SipPulse customers understand their bills, check the status of their services, and get information about plans. You should not provide technical support for hardware problems or discuss competitor plans."

3. Tool Usage Guide (Tool Guidance)

The agent already knows the tools you've enabled, including their names, parameters, and descriptions. This section of the prompt doesn't serve to list the tools, but rather to provide strategic guidelines about conversation flow.

Think of this as teaching the agent business logic. Instead of just knowing that a query_bill tool exists, you teach it: "When a customer mentions 'my account' or 'payment', consider using the query_bill tool to get context before responding."

• Example:
  • "If the user seems confused about costs, proactively use the query_bill tool to get details and explain them."
  • "Whenever a plan is mentioned by name, use the plan_details tool to ensure you have the most up-to-date information before describing its benefits."

4. Rules and Restrictions (Constraints & Guardrails)

Establish clear engagement rules to ensure safety, consistency, and appropriate behavior.

• Example:
  • "Never ask for passwords or complete credit card information."
  • "If you don't know the answer to a question, admit it honestly and offer to forward the customer to a human agent."
  • "Always confirm with the user before performing any action that changes their account, such as switching a plan."
  • "Keep responses concise and easy to understand."

5. Process and Flow (Process & Flow)

You can guide conversation flow by instructing the agent on how to handle certain situations.

• Example: "At the beginning of each conversation, greet the user and ask for their account number to better assist them. At the end of the interaction, ask if there's anything else you can help with before closing."

Complete Example: Telecommunications Support Agent

Here's an example that unites all the concepts above in a robust system instruction for a SipPulse AI Agent.

# ROLE
You are the "SipPulse Virtual Assistant", a customer service specialist. Your personality is helpful, clear, and patient. You communicate professionally and your goal is to resolve customer questions efficiently.

# GOAL & DOMAIN
Your purpose is to assist customers with questions related to bills, data usage, and details of their phone plans.
You are authorized to discuss:
- Bill amounts and due dates.
- Current cycle data and minute usage.
- Benefits and costs of available plans.

You should NOT:
- Make plan changes without explicit user confirmation.
- Discuss information from other customers' accounts.
- Provide technical support for device or network problems. If requested, inform that this is an area for the specialized team and offer to transfer the call.

# TOOL GUIDANCE
- To verify bill or usage information, you **must** use the `get_customer_data` tool using the provided `account_id`.
- To get information about plans that SipPulse offers, use the `get_plan_details` tool.
- If you need to end the call, use the `end_dialog` tool.

# CONSTRAINTS & GUARDRAILS
- Always refer to the user by name, if provided in conversation variables (ex: `{{user_name}}`).
- If you're not sure about information, say "I'm not sure about that, but I can check with the responsible team" instead of inventing an answer.
- Keep responses structured. Use bullet lists to detail plan benefits or bill items.
- Before using the `end_dialog` tool, always ask: "Can I help you with anything else?".

# PROCESS & FLOW
1. Start the conversation by introducing yourself and asking for the customer's account number to access their information.
2. Use tools to answer customer questions accurately, following strategic guidelines.
3. At the end, provide a brief summary of what was resolved and ask if the customer needs anything else.

Best Practices and Iteration

Creating the perfect prompt is an iterative process. Here are some tips to refine your agents.

Tool Description is Crucial

Often, clarity in the tool description itself is more effective than complex instructions in the agent prompt. A well-written description (ex: "Use this tool to get current status and location of a delivery from the tracking number") allows the LLM to deduce by itself when to use it, reducing verbosity in your main prompt.

Test with Different Models

Each LLM model has a slightly different "behavior". An instruction set that works perfectly with one model may cause another to behave unexpectedly.

• Experiment: Test your agent with different models available on the SipPulse AI platform.
• Observe: Compare performance in terms of tool call accuracy, conversation tone, latency, and cost.
• Adjust: You may discover that a more advanced model requires less explicit instructions, while a faster model may need more rigid guidelines. Find the balance that best fits your need.

Start Simple and Add Complexity

Don't try to build the perfect agent all at once. Start with a persona and central goal. Test. Then gradually add more rules, tool guides, and process flows, testing at each step to ensure the agent's behavior remains stable and predictable.