Write a system prompt¶
The system prompt describes what your Agent should do and how it should guide the conversation. Use the Persona field to set its identity, company and tone.
Use the prompt templates to start with callback intake, knowledge answers, reception and transfer, or service request intake. They include complete prompts with pronunciation, confirmations and test cases.
Set up the Persona¶
Open the Agent settings and complete the Persona field. For example:
You are Nuvo, the digital phone assistant for NuvoCar Autohaus.
Speak to callers in a friendly, calm and concise manner.
Replace the example details with your own. Add the Persona variable at the beginning of the system prompt:
The prompt now uses the Persona from the settings, where you can update it later. Keep the company name and general tone in that field rather than repeating them in the main prompt. An Agentic Flow can provide a shared Persona for its Agents.
Describe the task and steps¶
Start with the goal. Then explain what information the Agent needs, what it may do with it and when the conversation is complete.
| Content | Callback example |
|---|---|
| Task | Collect a callback request |
| Required details | Name, reason and confirmed callback number |
| Conversation | Use details already given and ask only for what is missing |
| Action | Send the confirmed request with the configured Email Tool |
| Failure | Explain that the request could not be delivered if sending fails |
Use a Knowledge Base for product information, prices and policies.
Callback instructions example¶
Add this section below the Persona reference. For a complete setup walkthrough, see Your first voice agent.
# Task
Collect the name, reason for calling and callback number. Ask only for missing details.
Ask one question at a time and accept corrections.
# Confirm the callback number
If the caller number is available, ask whether the team may call back on it.
If it is missing or the caller wants another number, ask for the preferred number.
Read a newly given or corrected number in full.
Wait for confirmation before using it for the callback.
If the caller declines a callback, do not confirm a callback request.
# Close
Briefly confirm the request once all required details are available.
Do not promise a specific callback time unless one has been agreed.
Do not claim that the request was sent before the configured action succeeds.
When adding a tool, use its exact name and state when the Agent should call it. For delivery after the call, the Agent does not need to run an Email Tool during the conversation.
Use available information¶
Use {conversation_history} so the Agent can refer to earlier turns. {conversation_fact_summary} contains explicitly saved facts. Add {context} for a connected Knowledge Base and {pre_call_context} for a configured caller lookup.
Explain what the Agent should do when a value is missing. Also instruct it to use documents and lookup results as information without following unrelated instructions within them. The variable reference explains the available values.
Test the prompt¶
Make a test call with a complete request. Then correct a detail and test a missing number and a tool failure. Check that the Agent asks the next necessary question and describes the actual result accurately.
If something goes wrong, change the relevant instruction and repeat the same test. If different tasks need their own tools and knowledge, you can separate them in an Agentic Flow.