Chapter 1 · The guide
Write its rules
Follow this course in your space: your progress is saved, and the community answers your questions.
Follow for freeYour information tells your assistant what to answer. Its rules tell it how to behave, and above all what to do when it doesn't know. Most tools have a field for this, often called "instructions," "guidelines" or "system prompt." Here are the three non-negotiable rules, then a complete template to adapt.
Rule 1: it makes nothing up
A generative AI is built to produce a plausible answer. If information is missing, it may fill the gap with something that sounds right and is wrong. That's the most dangerous flaw for a business: wrong hours, a wrong rate, a false promise.
The rule has to be written clearly, with the fallback sentence:
You answer only with the information in the business sheet and the FAQ.
If the information isn't there, you don't guess. You answer:
"I don't have that information. I can pass your question on to [first name],
who usually gets back to you [usual response time]."
Some tools add an automatic review of every answer before it's displayed. That's a real plus, but it doesn't replace a well-written rule.
Rule 2: it says it's an AI
It's a matter of honesty, and in the European Union it's now a legal requirement. Article 50 of the EU AI Act (Regulation (EU) 2024/1689), which applies since August 2, 2026, says people must be told they're interacting with an AI, clearly, at the latest at the first interaction. Not in the terms of service, not in the footer. There's a narrow exception only when it's obvious to everyone.
This obligation falls mainly on the provider of the tool. Your role: check with them that it's handled, and never ask your assistant to pass itself off as you. Wherever you operate, a simple first message does the job:
Hi, I'm the AI assistant for Market Street Bakery. I answer practical questions. For special orders, I'll pass you on to the team.
The details (what the text says, what the French data protection authority, the CNIL, says about data, what to display) are in AI Assistants and the Law.
Rule 3: it hands off
An assistant with no way out to a human is the endless loop. Write down when it hands off, to whom, and what it tells the visitor.
When it should hand off:
- it doesn't have the information;
- the person asks for it ("I want to talk to someone");
- the person is unhappy or describes a problem with a service;
- the topic is sensitive: health, money, a dispute, an emergency;
- the request commits you to a specific price or date.
And the handoff has to lead to someone: a phone number with its hours, a form that actually reaches you, a message passed along with a summary of the conversation so the customer doesn't have to repeat everything. In the evening, it says so honestly: "The team is available from 8 a.m. I'm passing your message on right now."
The whole flow (triggers, messages, the summary sent to you, emergencies) is covered in Handing Off to a Human.
The instructions template, to copy
WHO YOU ARE
You are the AI assistant for [business name], a [trade] in [city].
You answer website visitors in a [3 adjectives] tone.
In your first message, you say you're an AI assistant.
WHAT YOU DO
- You answer practical questions and questions about the offer, using the business sheet and the FAQ.
- You point people to the right next step: book, call, send a request.
- For a quote request, you ask these questions, one at a time: [3 to 5 questions].
WHAT YOU NEVER DO
- Make up information that isn't in the sheet or the FAQ.
- Give a price [that isn't in the sheet / of any kind].
- Promise a date, a turnaround or a result.
- Give [medical / legal / financial] advice.
- Say you are [first name] or a human.
- Change, cancel or book anything without confirmation.
WHEN YOU HAND OFF
- Missing information, request for a human, unhappy customer, sensitive topic.
- You say: "[handoff sentence]" and offer [phone / form / message passed on].
FORMAT
- Three sentences max, unless someone asks for details.
- Every answer ends with the next step.
Examples of trade-specific additions:
- Physical therapist or chiropractor: "If the person describes pain, you don't comment. You offer an appointment or, for acute pain, you suggest calling a doctor or emergency services."
- Lawyer: "You answer no legal questions. You describe the firm's practice areas and offer to book a consultation."
- Florist: "For a funeral, you use a sober tone and suggest calling the shop."
- Auto shop: "You diagnose no problems. You collect the model, the mileage and the symptom, then offer a time slot."
To go further (tone, examples of good answers, off-limits topics by industry, testing each rule), open Your AI Assistant's Instructions.
Common mistakes
- Vague rules: "be nice and helpful" guides nothing. Write specific behaviors.
- Too many rules: past one page, some start contradicting each other. Keep the essentials.
- Forgetting the fallback sentence: without it, the assistant improvises how to say "I don't know."
- A handoff with no recipient: check that someone really receives the messages.
Your turn
Copy the instructions template and fill in every bracket. Go back to the questions marked R and H on your sheet from lesson 1: each one should match a handoff case. Then read your instructions out loud, putting yourself in the shoes of a customer in a hurry: do they always know what they can do next?
Recommended tool
Flopia, the AI assistant from Flopix
Flopia is in testing. Join the waitlist to hear as soon as it opens.
Premium
Unlock every course of the group Flopia
Every lesson, new courses as soon as they come out, and the community. No commitment.