How to Write Instructions for Your AI Support Agent (Tone, Rules, Boundaries)
Published on by
Antoni
Two agents can read the same knowledge base and still feel completely different to a customer. One sounds like a colleague who knows the product and the company's manners. The other sounds like the Bot: technically correct, oddly stiff, and lost the moment a question wanders off the page. The difference is rarely the model. It is the instructions. Here is what belongs in them, what belongs in the knowledge base instead, a template, and how to test and revise them.
Key Takeaways
- Instructions set how your agent talks and behaves; the knowledge base holds what it knows. Keep facts out of the instructions.
- Cover seven things: who you are, who your customers are, tone with examples, formatting, hard rules, when to hand off, and how to handle unknowns.
- Specific beats vague. "Always mention the 30-day guarantee on refund questions" changes replies; "be helpful about refunds" changes nothing.
- Test every change against ten real past questions, and revisit the instructions monthly against handed-off conversations.
Instructions are where "sounds like us" comes from
The model gives your agent language. The knowledge base gives it facts. Neither tells it that you sign off with first names, never promise a delivery date, or take a customer who says "cancel" seriously rather than talking them out of it. That layer is the instructions, and most teams either skip it or stuff it with policy nobody tests.
In FutureBase this layer lives on your agent's Settings page: a Tone of voice preset (friendly, professional, supportive and so on) and a free-text Custom instructions field the agent follows in every reply, on every channel. Pick the preset closest to your brand, then use the instructions for everything the preset can't know. The customizing your agent guide covers both.
One rule before anything else: your agent still answers only from your knowledge. Instructions shape behaviour; they do not add facts. If "do you ship to Canada?" is not in your knowledge base, no instruction will conjure the answer.
What to put in
Think of the instructions as the day-one note for a new support hire, minus the product details they will find in the help center.
1. Who the company is
Two or three sentences: what you sell, to whom, and the one thing a new hire should understand about how you treat customers. An agent that knows it serves freelancers makes different small choices than one that thinks it serves enterprise security teams.
2. Who the customers are
Who is on the other side of the chat, what they usually want, and how they usually feel. "Mostly non-technical" changes how much jargon is acceptable; "often messaging from a phone" argues for short answers.
3. Tone, with examples
The preset sets the register; the instructions set the specifics. Adjectives ("warm, professional, concise") help a little; examples help far more. Give the agent a few pairs: a phrasing you would never use and the one you want instead.
| Instead of | Say |
|---|---|
| "I apologize for any inconvenience this may have caused." | "Sorry about that, let's sort it out." |
| "Please be advised that your request has been received." | "Got it. I've passed this to the team." |
Spell out the small things too: first names, emoji, "we" or "I".
4. Formatting rules
Say how long a reply should be, when a numbered list is appropriate (steps), and whether to link to help pages. "Lead with the answer in one sentence, then add detail only if the question needs it" fixes more bad replies than any other single line.
5. What it must never do
The boundary section. Give it its own heading so it is easy to audit:
- Never promise a refund, credit, delivery date or feature date; hand off if the customer needs an exception.
- Never give legal, medical, tax or financial advice.
- Never ask for passwords, full card numbers or other secrets.
- Never claim the company does something the knowledge base does not say it does.
Keep it short and absolute; "try to avoid" gets treated as a suggestion.
6. When to hand off
Hand off too late and you frustrate people; too early and the agent is just a slower contact form. Name the triggers: the customer asks for a person; the question is a refund exception, a billing dispute, a security or legal matter; the customer is going in circles or upset. Say what to do at that moment too: acknowledge, summarise what was tried, set an expectation.
In FutureBase the hand-off itself is configured on Actions → Hand off to your team: who hears about it (Live Desk, email, Slack or Discord) and the first message the customer reads when a person takes over. The instructions decide when the agent reaches for it.
7. How to handle unknowns
This is what separates an agent from the Bot. The Bot says "I didn't quite get that" and offers a menu. Tell your agent to say plainly that it does not have the answer, not to guess, and to offer a person.
What to leave out
The most common failure is an instructions field that has quietly become a second, worse knowledge base. Prices, plan limits, shipping times, the refund window: each belongs in the knowledge base as an FAQ, a snippet or a page, where it can be found by search, kept in sync with your site, and flagged when it contradicts something else. Written into instructions, a fact is invisible to all of that and goes stale silently: the pricing page updates, the instruction does not, and the agent quotes last year's price.
| Belongs in custom instructions | Belongs in the knowledge base |
|---|---|
| How to greet and sign off | Opening hours and contact details |
| "Always mention the 30-day guarantee on refund questions" | What the guarantee actually covers |
| "Never promise a delivery date" | Shipping times by region |
| Tone examples and formatting rules | Plan features, limits and prices |
A good test for any sentence: if it could change when marketing updates the website, it is a fact and it moves out. If it would still be true after a full rebrand, it is behaviour and it stays. For the facts side, see how to train AI on your documentation. And keep it under a page; every rule narrows judgment elsewhere.
A fill-in-the-blanks template
Paste this into Custom instructions, replace the brackets, and delete any section you do not need.
ABOUT US
[Company] makes [product] for [who]. [How we treat customers, in one line.]
OUR CUSTOMERS
Mostly [role / technical level], usually asking about [top three things].
TONE
Sound like [a knowledgeable colleague]. Emoji: [never / sparingly].
Instead of "[stiff phrase]", say "[our phrase]".
FORMAT
Lead with the answer in one sentence. Keep replies to [2-4] sentences.
Numbered lists only for steps.
ALWAYS
- When a customer asks about [topic], mention [policy to surface].
NEVER
- Promise a refund, credit or delivery date.
- Give legal, medical, tax or financial advice.
- Claim we do something our knowledge does not say we do.
HAND OFF TO A PERSON WHEN
- The customer asks for a person, or is upset.
- The question is about [refund exceptions / billing disputes / security].
When you hand off, summarise what you tried and say [what happens next].
WHEN YOU DON'T KNOW
Say so plainly. Do not guess. Offer a person.
Before and after: a weak instruction and a strong one
The usual starting point:
Be friendly and helpful. Answer customer questions about our product.
Escalate to a human if you can't help. Be professional.
Every clause is something the agent would do anyway. "Friendly" and "professional" pull in opposite directions with no examples to resolve them, and "if you can't help" leaves the hand-off threshold to the model. The same intent, written so it changes replies:
Northwind is a scheduling tool for independent fitness coaches. Most
customers run a one-person business and message us from their phone.
Sound like a calm, knowledgeable colleague, not a help desk. Lead with the
answer and keep it to three sentences.
When anyone asks about refunds, mention the 14-day guarantee. Never
promise a refund outside it; hand off instead.
Hand off when the customer asks for a person, when a payment went wrong,
or when the customer is frustrated. Say what you tried and that a
teammate will reply here.
If the answer is not in your knowledge, say so and offer a person.
Barely longer, but every sentence changes a specific reply. On "I was charged twice, what's going on?" the weak version produces a generic apology; the strong one apologises briefly, does not speculate about the charge, and hands off with the context in place.
Test changes against real past questions
Never judge an instruction by the questions you think customers ask. Use the ones they asked.
- Build a test set. Open your closed conversations (in FutureBase, the Closed view in Live Desk) and pull ten to fifteen real questions: routine ones, a few that were handed off, one or two from annoyed customers, and at least one the agent got wrong. Reuse the set every time.
- Change one thing at a time. Changes apply to new replies straight away, so you can edit, test and edit again in minutes. Change two things at once and you will not know which one fixed, or broke, a reply.
- Ask the set again. Ask your agent on the Knowledge page shows the answer and the knowledge it used, so you can tell an instructions problem from a knowledge gap. The shareable chat link (Channels → Share link) shows the full conversation, including hand-off wording.
- Read the answers as a customer. Did it lead with the answer? Did it hand off at the right moment, with the right words? Did it make something up where it should have admitted a gap?
| What you see | Where the fix lives |
|---|---|
| Right facts, wrong tone or length | Custom instructions |
| Wrong or missing facts | Knowledge base (FAQ, snippet or page) |
| Handed off too early or too late | The hand-off triggers in your instructions |
Iterate monthly
Instructions are not a launch artifact; the worries of month one are rarely the ones that matter in month six. Book thirty minutes a month:
- Read every handed-off conversation. Live Desk shows why each one was handed off. Should the agent have handled it (a knowledge gap, or a too-eager trigger), or was the hand-off right but worded poorly?
- Read ten random resolved conversations. Resolved is not the same as good. Look for replies that were correct but long, cold, or missing a policy you always want mentioned.
- Add at most one or two rules, remove any that no longer earn their place, and re-run your test set. A never-list that grows monthly means facts are leaking in.
Over a few months the instructions get shorter and sharper, the knowledge base gets fuller, and the agent starts sounding like someone who works there.
Frequently asked questions
How long should AI support agent instructions be?
Shorter than you expect: usually half a page to a page. Instructions describe behaviour, not facts, and every extra rule narrows the agent's judgment elsewhere. Start with a short template and add a rule only when a real conversation shows you need it.
Should I put prices and policies in the instructions or the knowledge base?
In the knowledge base. Instructions should say how to talk about a policy, for example "always mention the 30-day guarantee on refund questions", while the knowledge base holds what the policy actually says. Facts written into instructions go stale silently and are invisible to knowledge search and conflict checks.
How do I test a change to my agent's instructions?
Keep a set of ten to fifteen real past customer questions, including some that were handed off and some that went wrong. Change one thing at a time, ask the set again, and read the answers as a customer would. If the facts are wrong, fix the knowledge base; if the tone, length or hand-off timing is wrong, fix the instructions.
How often should I revise the instructions?
Monthly is a good default. Read the month's handed-off conversations and a sample of resolved ones, add or remove at most a rule or two, re-run your test set, and move any fact that has crept into the instructions back to the knowledge base.
Want an agent that sounds like your team from the first reply? Start free with FutureBase, paste in the template above and test it against your real questions. See how teams use it for SaaS customer support and 24/7 customer support.
