VibotGetting Started3. Configure the agent

3. Configure the agent

View as Markdown

Open AI → AI Agent in the sidebar. This page defines who the AI agent is and how it should behave.

Persona, language, system prompt, guardrails — the fields you'll touch most.

How the page is laid out

Persona name sits in the header, above everything else — it applies no matter which tab you’re on. The rest is split across six tabs:

TabWhat’s in it
ConfigureResponse language, voice replies
PersonaSystem prompt, initial greeting, guardrail rules
VariablesThe information the agent collects from customers
ToolsThe actions the agent can take mid-chat
AnalysisPost-conversation sentiment and goal scoring
AdvancedReply footer and Vibot branding

Save changes applies the whole page, not just the tab you’re looking at.

Persona name

The name the agent uses when greeting customers (e.g. Priya, Arjun, Maya, Support). Keep it short and consistent with your brand voice. Max 100 characters. Lives in the page header.

Persona → System prompt

The core instructions — max 10,000 characters.

A good system prompt:

  • States clearly who the agent is and what business it represents.
  • Describes the tone — formal, friendly, concise, etc.
  • Lists what topics the agent should and should not discuss.
⚠️

Do not put product details, prices, or FAQ content in the system prompt. Those belong in the Knowledge Base (next step) — the agent retrieves them automatically when relevant.

Example:

Text
You are Priya, a customer support assistant for Acme Electronics.
You help customers with product questions, order status, and returns.
Be friendly, concise, and professional.
Always offer to connect the customer with a human agent if you cannot resolve their issue.
Do not discuss competitors or make promises about delivery timelines.

Click the purple ✨ Generate with AI button above the textarea to open the System Prompt Generator wizard in the Vibot Assistant — it takes your free-form description and emits both a polished prompt and an opening greeting line.

Persona → Initial greeting (optional)

A short opening message used when starting a brand-new conversation. 600-character cap. Leave empty to fall back to the agent’s natural opener.

Persona → Guardrail rules

Hard constraints appended to every system prompt. Max 5,000 characters.

Examples:

Text
Never share internal pricing structures with customers.
Do not provide medical, legal, or financial advice.
Always escalate complaints about damaged goods to the support team.
Never discuss competitor products by name.

Optional, but strongly recommended.

Configure → Response language

SettingBehaviour
Auto-detect (recommended)Agent detects the customer’s language per message and replies in the same language.
Locked languageAgent always replies in the chosen language regardless of what the customer writes.

Available languages: English, Hindi, Marathi, Gujarati, Tamil, Telugu, Kannada, Bengali.

Configure → Voice replies

When enabled, the agent transcribes inbound voice notes (Whisper) and replies in audio.

PlanVoice replies
Free, LiteLocked off — upgrade required
BasicEnabled, 20 replies per billing cycle (counter resets on subscription renewal)
Pro and aboveUnlimited

The Voice replies card shows your usage for the current cycle when you’re on Basic.

Variables

Variables are the specific fields you want every conversation to produce — a name, a budget, a preferred date. Declare them here and the agent gathers them in conversation, at the right moment, instead of interrogating the customer up front.

You don’t wire anything up. Declaring a variable is enough: the agent is given a capture_variable action automatically and records each field as the customer provides it.

Each variable has four parts:

FieldMeaning
KeyThe stable identifier, lowercase with underscores (customer_name). Captured values are stored under it — keep it stable once conversations are using it.
DescriptionWhat to collect, in plain language (“customer’s full name”). The agent reads this.
Typetext, number, date, email, phone, or boolean.
RequiredWhether the agent should make a point of gathering it.

Up to 20 variables. Captured values appear on the conversation in Chats and travel with the contact.

phone-typed variables are never asked for. WhatsApp already tells us the number the customer is messaging from, so Vibot fills any phone variable from that identity rather than reading it out of the chat. This stops the agent asking a customer for the number it is currently messaging them on — and stops it mistaking a number it mentioned for the customer’s own.

The agent is also told what it already knows about a returning customer, so it won’t ask twice for something captured in an earlier conversation.

Tools

The actions the agent can take mid-conversation — capturing a lead, escalating to a human, booking through a connected provider, or calling your own endpoint. This tab is a shortcut into the full Tools workspace.

Analysis

Off by default. When enabled, a second AI pass reads each conversation once it goes quiet and scores it — how the customer felt, whether your goal was met, and what to do next. Results appear on the chat and on your Overview.

Success criteria is where you say what a good conversation looks like, in one or two plain lines (max 500 characters) — for example “Customer shared their budget and agreed to a site visit.” The AI reads it verbatim and marks each conversation met or not met against it.

⚠️

Analysis costs one extra AI call per conversation, billed like any other message against your plan quota. That’s why it ships off.

Advanced → Vibot AI badge

Adds “Powered by Vibot AI” to every reply. On Basic and above it’s optional, and leaving it on takes 10% off your next billing cycle. On Free it is a mandatory reply footer and cannot be switched off.

Saving

Click Save changes. A green checkmark confirms the save. Changes take effect on the next incoming message — no restart, no downtime.

Next step

Send your first message to verify everything works → Your first conversation.