Blog · · 3 min read
System prompt for a customer-service AI agent: free template and how to fill it in
A free system prompt template for a customer-service AI agent (WhatsApp, web chat or email), with the sections that matter, examples and the mistakes that make bots make things up.
The difference between an AI agent that actually helps your customers and one that invents prices is rarely the model. It's the system prompt: the fixed text the agent reads before every conversation that tells it who it is, what it knows and what it must never do.
Here's a template you can copy, what each block does and the most common mistakes.
The template
You are the virtual assistant for [BUSINESS NAME], [what the business does in one sentence].
You help customers via [WhatsApp / web chat / email] in [language(s)].
GOAL
Answer questions about [services/products], opening hours and bookings, and hand off
anything you can't resolve to a person.
BUSINESS INFORMATION (single source of truth)
- Opening hours: [ ... ]
- Address / service area: [ ... ]
- Services and prices: [ ... ]
- How to book or buy: [ ... ]
- FAQ: [ ... ]
RULES
1. Use ONLY the information above. If it isn't there, say you don't know and offer
to connect the customer with a person. Never invent prices, timelines or availability.
2. At the start, make clear you are an automated assistant.
3. Keep replies short: 3 sentences max, [friendly / formal] tone, no jargon.
4. If the customer types HUMAN, is upset or asks for something out of scope,
reply: "I'll pass you to someone on the team, they'll message you shortly."
5. Never ask for sensitive data (cards, passwords, ID documents).
6. Reply in the language the customer writes in.
What each block does
- Identity and channel. Gives context: a WhatsApp assistant (short messages) shouldn't sound like an email one.
- Goal. Sets the boundaries. An agent without a goal tries to help with everything, including things it shouldn't.
- Business information. The most work for you and the most value. Write real prices and hours exactly as you'd tell a customer.
- Rules. Rule 1 matters most: it forces the agent to admit what it doesn't know instead of filling gaps with something that sounds right.
The 5 most common mistakes
- Vague information. "Affordable prices" doesn't help: when a customer asks how much, the agent will improvise. Give numbers, or say explicitly that pricing is quoted in person.
- No instruction for "I don't know". Without a clear way out (hand off to a person), the model tends to make things up.
- Huge prompts. Pasting your whole website adds noise. Summarize what people actually ask.
- Not disclosing it's a bot. Besides being honest, it avoids confusion and complaints.
- Not testing it. Before going live, ask it 20 real questions, including trick ones ("can I get 50% off?"), and fix the prompt.
Test it in 10 minutes
- Open your agent in n8n (or your tool's test chat) and paste the prompt.
- Write a list: your 10 most frequent questions and 10 it shouldn't be able to answer.
- Mark the wrong answers and add the missing information or rule.
- Repeat until all 20 pass.
Honest limitations
- A good prompt greatly reduces made-up answers, but doesn't eliminate them completely. Review real conversations for the first few weeks.
- If your information changes often (stock, availability), have the agent look it up in a spreadsheet or calendar rather than hard-coding it in the prompt.
Next step
This template is already built into our free WhatsApp agent for n8n, with per-customer memory and human handoff. If you don't have n8n yet, start with how to install n8n for free. And if you want all 8 agents ready to go, check out the AgentKiosk pack.
