Skip to main content
The persona is what an Advanced agent is on every step: who it is, what it is for, the rules it works by, how it handles pushback, and what it never does. The flow says what to do at each step; the persona says how. Open it from the left rail: Agent Persona. Each part has its own tab, and each tab uses the whole page.

​
Role

One or two short paragraphs. Name the agent, the business, the channel and the tone.
You are Riya, the WhatsApp assistant for Glow Clinic, a skin and hair clinic in Bengaluru. You are warm, brief and professional. You are a virtual assistant and say so if asked.

​
Goal

What the flow is driving towards, in one or two sentences. The AI uses it to decide what to do when the flow has nothing planned — after an end step, or after a side question.
Book a free skin consultation for every interested lead, and answer their questions about the clinic along the way.

​
Instructions

Rules that hold on every step. Put here what you would otherwise repeat in every step:
  • Language — which languages to reply in, and how to detect the user’s.
  • Approved facts — prices, opening hours, what the offer includes. The AI states only what is written down.
  • Formatting — message length, emojis, how many questions per message.
  • What to do off-flow — how to treat a question the flow does not cover.
Keep step-specific text in the step itself: an instruction that only matters at one point of the conversation belongs in that step’s message or guidance.

​
Rebuttals

Each row is an objection (what the user might say) and a response (how to answer). The AI puts the response in its own words, in the conversation’s language, then returns to the step it was on. Write the trigger the way users actually say it — add examples in brackets ("mehnga hai", "baad mein") when the chat is multilingual.

​
Guardrails

The Rules list is the agent’s hard limits: never promise results, never share another customer’s details, never type raw function output. They apply above everything else. Above the list are the two snippets every chat agent carries:
  • Security guardrails — how to refuse questions about internal systems.
  • Function calling — how to talk while a function runs, and to never show raw results.
Leave them as they are unless you have a reason; an empty snippet falls back to the default text.

​
Reorder by dragging

Instructions, rebuttals and guardrails are read top to bottom, so order matters: put the rule that wins a conflict first. Each row has a drag handle on the right, above its delete button — drag it to move the row. A row can also be picked up from the keyboard: focus the handle, press Space, move with the arrow keys, and press Space again.

​
Where it ends up

When the agent loads, the persona and the flow are compiled into one system prompt in this order: role, goal, instructions, the flow’s steps, global intents, objection handling, then guardrails and the two snippets. The compiled prompt is built on the server and is not shown in the dashboard.