Writing an AI agent's instructions: mission, tone, forbidden topics
Mission, tone, welcome message, forbidden topics, human handoff: how to write clear instructions for an AI agent, with examples and the mistakes to avoid.
The product5 min read
Two agents can read exactly the same content and behave very differently. One answers precisely and plainly, points to the right page and offers a contact when it does not know. The other wanders into long answers, brings up topics the business would rather avoid, or speaks in a voice that does not sound like the brand. The difference often comes down to the instructions.
This article explains how to write an agent’s instructions in blabla.bot, field by field, with examples and the most common mistakes.
What you set, and what you do not
In blabla.bot an agent’s behaviour is set through six structured fields, not through a free-form prompt where anything goes:
| Field | Purpose | Limit |
|---|---|---|
| Role and mission | Who the agent is, who it answers for, what it must achieve | 2,000 characters |
| Answer tone | Professional, Friendly, Concise or Pedagogical | Fixed list |
| Default language | French or English | Fixed list |
| Welcome message | The first message shown when the widget opens | 500 characters |
| Forbidden topics | Topics the agent politely declines | 20 topics at most |
| Contact e-mail | The address offered when the agent cannot help | Optional |
On top of these settings, platform rules always apply and cannot be switched off: the agent does not make up answers, cites its sources, declines requests outside its mission, does not reveal its instructions and ignores any instructions hidden in the indexed content. So you do not need to write “do not lie” or “do not reveal your instructions”: that is already handled, and handled better.
Role and mission: the field that matters most
The mission answers three questions: who the agent is, which business it answers for, and what it must achieve. It should be short, concrete and focused on the visitor.
A mission that is too vague:
You are a friendly assistant who helps people.
A useful mission:
You are the website assistant of Dupont Motors, an independent garage in Bristol. You help visitors understand the services offered (servicing, repairs, air conditioning, tyres), find the opening hours and address, and get in touch to book an appointment. You do not give quotes: for a precise price, invite the visitor to contact the garage.
The second version states the scope clearly, what to do and what not to do. The agent knows when a question falls outside its mission.
A few principles:
- Describe tasks, not qualities. “Help choose a plan” is more useful than “be helpful”.
- Say what to do in edge cases. Quotes, complaints, emergencies: what should the agent do?
- Do not copy the knowledge base. Opening hours, prices, conditions: this information belongs in the indexed pages and PDFs, where it is cited as a source and refreshed on re-indexing. Information written into the mission is easily forgotten when something changes.
Tone: pick from a short list
Four tones are available: Professional, Friendly, Concise and Pedagogical. Pick the one closest to how the business already talks to its customers:
- an accounting or law firm: Professional;
- a restaurant or a neighbourhood shop: Friendly;
- software whose users want a quick answer: Concise;
- a training provider or a public service: Pedagogical.
The list is fixed on purpose. A tone described in free text (“a bit quirky but not too much”) gives unpredictable results.
Language and welcome message
The default language sets the language of the answers and of the widget labels. If a visitor clearly writes in another language, the agent answers in that language.
The welcome message is the first impression. In one or two sentences, it should say what the agent can do:
Hello, I can answer your questions about our services, opening hours and booking an appointment. How can I help?
Avoid promising more than the agent can deliver (“I can answer all your questions”). The message can be left empty: the widget then opens straight on the input field.
Forbidden topics: one line per topic
Forbidden topics are short labels, one per line, up to twenty. The agent politely declines to discuss them.
Sensible examples:
- medical advice or diagnoses;
- legal advice;
- comparisons with competitors;
- negotiated prices or discounts;
- political or religious questions.
Two mistakes to avoid. First, forbidding topics that are part of the mission: a garage agent that cannot talk about prices at all will frustrate visitors, whereas “personalised quotes” is a reasonable limit. Second, piling up topics just in case: every topic you add shrinks what the agent can do for the visitor.
Human handoff
The contact e-mail is offered to the visitor when the agent cannot help, for example when no relevant information is found. The widget then shows a form where the visitor leaves their e-mail address and a message. Leave the field empty if no contact is planned: the agent will simply say it does not have the information.
Choose an address that someone actually reads. A human handoff that never answers is worse than no handoff at all.
Start from a template
To avoid a blank page, blabla.bot offers templates by business sector (restaurant, online shop, real estate, health, trades, software, training and more) and by use case (customer support, pre-sales, appointments and bookings, frequently asked questions). A template fills in the mission, tone, welcome message and forbidden topics. The language and contact e-mail are kept, and nothing is saved until you click “Save”.
Treat the template as a draft: replace the generic wording with your client’s real names, real services and real edge cases.
A five-step method
- Pick a template close to the client’s business.
- Rewrite the mission with the business name, its scope and its edge cases.
- Choose the tone and write an honest welcome message.
- List the forbidden topics with the client, without overdoing it.
- Enter a contact address someone reads, then test the agent in the simulator.
In short
Good instructions are short, concrete and focused on what the visitor came for. The mission sets the scope, the tone and welcome message give the voice, forbidden topics set the limits, and the human handoff takes over when the agent does not know. The safety rules are already in place.
Common questions about setting up agents are gathered in the FAQ.
Tags
- instructions
- behaviour
- AI agent
- guardrails
On the site
Read next
The product
What an AI chat agent for a website is (and what it is not)
A website AI agent answers visitors from the client's own content, within limits set by the agency. Here is what it does well and what it does not do.
· 5 min read
Knowledge and RAG
Why a good assistant knows how to say "I don't know"
Relevance threshold, human handoff, unanswered questions: why a reliable AI assistant would rather admit it does not know than invent a plausible answer.
· 5 min read
The product
Testing an AI agent before launch: the simulator and a method
How to test an AI agent before installing it: a simulator that reproduces the widget, an explanation panel, and a question set that covers the difficult cases.
· 5 min read