Prompt and initial message

Open your application in the dashboard, on its Application Data tab. Two fields there shape how the agent behaves and how a conversation starts.

System Prompt

The System Prompt is the agent's instructions: who it is, what it should help with, what tone to use, and what it should never do. It is sent with every message in the conversation, so it stays in effect for the whole session, not just the first reply. Editing it takes effect on the very next message, nothing needs to be redeployed.

A prompt that works well usually covers:

• Role and scope — what the agent is for, and what is out of scope.
• Tone — formal, casual, how it should refer to itself and the user.
• Tool usage — when it should look things up in a connected API or the knowledge base instead of answering from memory, and what to do when a lookup fails.
• Boundaries — topics or actions it should decline, and how to respond when it does.

You are the support assistant for Acme's billing portal.
Answer questions about invoices, subscriptions and payment methods.
Always look up real data through the connected tools before answering;
never guess an amount, a date or an order status.
If a request needs a human (refunds above 500 CHF, account deletion),
say so plainly and do not attempt it yourself.
Keep answers short and in the same language the user wrote in.

Initial Message

The Initial Message is the first thing the user sees, before they type anything. It is returned in the initialMessage field when a session is created, so your interface can show it immediately without waiting on a model response.

Hi! I can help with invoices, subscriptions and payments. What do you need?

Both fields are saved with the application's Update button.