Agent Setup

What is the system prompt?

The system prompt is a set of instructions you write that tells your AI agent how to behave. Think of it as a detailed briefing for an incredibly capable employee — one that can browse the web, run code, generate images, automate websites, collect payments, and connect to any service you need.

Your system prompt shapes everything: the agent's personality, its decision-making, when it uses which tools, how it talks to people, and what it refuses to do. The better your instructions, the more effectively the agent uses its full range of capabilities.

Every time someone messages your WhatsApp number, the AI reads your system prompt before responding. This is how it knows your business, your rules, and how you want things done.

Writing your system prompt

Go to your instance settings and find the System Prompt text area under Core Configuration.

Tips for a good system prompt

Be specific about identity. Tell the agent who it is and what it represents.

You are the AI assistant for Swift Deliveries, a courier service
operating in Nairobi, Mombasa, and Kisumu. You handle customer
inquiries, track shipments, process payments, and manage bookings.

Define what it should do — and think big. The agent can do far more than just chat. Tell it to use its capabilities.

Your responsibilities:
- Answer customer questions using our knowledge base
- Look up real-time tracking information by browsing our tracking portal
- Generate delivery quotes by running calculations
- Create custom route maps when customers need visual directions
- Collect payments via M-Pesa when orders are confirmed
- Schedule follow-up messages to check on delivery satisfaction
- Research competitor pricing when the sales team asks

Tell it when to use specific tools. The agent has many tools available — guide it on which ones to use for different situations.

When a customer asks about a competitor's service:
- Use web browsing to look up the latest information
- Don't rely on old knowledge — always check the live website
 
When a customer needs a data report:
- Create a code sandbox and write a Python script to process the data
- Generate charts if they'd be helpful
- Export the result as a file and send it back
 
When a customer wants a custom image:
- Generate it using the image tools
- Ask for feedback and iterate if needed

Set boundaries. Tell the agent what it should NOT do.

Do not:
- Make promises about refunds without checking with management
- Share personal information about staff or other customers
- Log into customer accounts without explicit permission
- Send bulk messages without approval

Describe the tone. How should the agent communicate?

Be warm, professional, and efficient. Use simple language.
Keep messages concise — most people are reading on their phones.
You can use Swahili greetings when appropriate.
Don't use emojis excessively.

Give it fallback instructions. What should it do when stuck?

If you can't find the answer in the knowledge base or on the web,
say: "Let me connect you with our team for that."
Then escalate the conversation to a human agent.

Example: Full system prompt for a versatile agent

You are the AI assistant for Apex Solutions, a technology consulting
firm in Nairobi.
 
You have powerful tools at your disposal — use them proactively:
 
INFORMATION:
- Search our knowledge base for company policies, pricing, and services
- Browse the web for any real-time information you need
- Read competitor websites, industry news, or documentation as needed
 
ACTIONS:
- Collect payments from clients via M-Pesa for invoices and retainers
- Schedule follow-up messages and reminders for the team
- Generate reports by writing and running code in sandboxes
- Create presentation visuals and diagrams using image generation
- Automate data entry on our project management portal via browser
 
COMMUNICATION:
- Be professional but approachable
- Use clear, concise language
- Respond in the language the client uses (English or Swahili)
- When sharing data or lists, use bullet points
 
BOUNDARIES:
- Never share confidential client information across projects
- Always confirm payment amounts before sending STK push
- For contracts over KES 500,000, escalate to a senior consultant
- If you're uncertain about technical advice, say so and offer to
  connect the client with a specialist

Saving changes

After writing or editing your system prompt, click Save Changes. The new instructions take effect immediately — the very next message your agent receives will use the updated prompt.

You don't need to restart or reconnect anything.

Reference images

You can upload images that your AI agent can reference when responding to customers. This is useful for things like product photos, menus, price lists, maps, or diagrams.

Adding reference images

1. Scroll to the Reference Images section below the system prompt
2. Click Add Images
3. Select image files (PNG, JPG, GIF, or WebP — maximum 10 images per instance)
4. For each image, fill in:
   • Description (required) — What the image shows (e.g., "Our full lunch menu with prices for 2024")
   • Usage hint (optional) — When the AI should reference this image (e.g., "Show this when customers ask about lunch options or prices")
5. Click Save

The AI uses these images as visual context. If you upload a product catalog image with clear descriptions, the agent can reference specific items, prices, and details from it when customers ask.

Managing images

• Click the edit icon on any image to update its description or usage hint
• Click the delete icon to remove an image
• You can have up to 10 reference images per instance

The agent learns on its own

Your agent doesn't just follow static instructions — it builds up knowledge over time:

• Skills — When the agent figures out how to do something (like a multi-step browser workflow or a data processing task), it can save that as a reusable skill. Next time a similar request comes in, it already knows how.
• Browser automations — When the agent successfully automates a task in its browser (like logging into a portal and extracting data), it saves the workflow. Future runs are faster and more reliable.
• Notes — The agent can save working notes about ongoing tasks, customer preferences, or context it needs to remember across conversations.

You don't need to configure any of this — it happens automatically as the agent works.

Next steps

• Auto-Responder Rules — Control who your agent responds to and when
• Knowledge Base — Upload documents for the agent to reference
• AI Providers & Models — Choose which AI model powers your agent
• Integrations — Enable code sandboxes, image generation, payments, and more