An assistant with no persona is not neutral. It is a stranger with an opinion about how your business talks — cheerful, hedging, faintly American, allergic to a straight answer. Nobody chose that. It is just what you get when the field is blank.
This post is the practical version: which controls actually shape how the assistant sounds, in the order you should touch them, and how to tell whether a change worked. If you want the theory behind prompt design — structure, common failure modes, a template — read System Prompt Engineering alongside this. This one is about the buttons.
Where the persona lives
Open Assistants, click your assistant, and stay on the General tab. Everything in this section is on that tab.
Four controls do the work, and they do different jobs:
| Control | Section | What it shapes |
|---|---|---|
| Instructions | AI Behavior | Who the assistant is, how it speaks, what it refuses |
| Goal | Outcomes | What it should be steering the conversation toward |
| Suggested Questions | Chat | What visitors think it can do, before they type anything |
| Temperature | AI Model | How varied its wording is from turn to turn |
Everything is held locally until you press Save (⌘S / Ctrl+S). An Unsaved changes marker appears in the header while you have edits outstanding, and Preview always opens the last saved version — so save before you go looking for your change.
1. Write the first line before anything else
The opening line of Instructions frames every line after it. Give it a role and an audience, in that order:
You are the front-desk assistant for Northgate Veterinary. You talk to pet owners — existing clients and people deciding whether to register.
That single sentence quietly resolves a dozen ambiguities: register vs. book, "we" vs. "they", how much jargon is acceptable, whether "your dog" is presumptuous. Compare it to "You are a helpful assistant for our company," which resolves none of them.
Instructions holds up to 10,000 characters. A counter under the box shows how much you have used and turns amber once you pass 8,000. You will not get near it, and you should not try.
2. Describe behaviour, not adjectives
"Friendly" is not an instruction. It is a hope. Models follow described behaviour far more reliably than personality labels, so write the behaviour down:
Vague: Be friendly and professional.
Followable: Greet by first name when you know it. Answer in two or three sentences. Once a question is answered, offer the next useful step rather than asking "is there anything else?"
The same applies to formality. Instead of "be casual," give it a swap it can copy: "Say 'you're all set' rather than 'your request has been processed.'" One concrete example beats three adjectives.
Five behaviours worth spelling out explicitly, because the default is rarely what you want:
- Length. "Keep answers under 60 words unless asked for detail" is a good default.
- Uncertainty. What should it say when it does not know? Say the words you want.
- Names and pronouns. "We" or the business name? Both, but which by default?
- Repetition. Whether it may re-ask for something the visitor already gave.
- Openers. Ban "Great question!" now, or read it four hundred times later.
3. Say what is out of scope
Most bad answers are not wrong answers. They are answers to questions the assistant should have declined.
Name the off-limits topics, and give the exact redirect you want:
Do not quote prices, discuss insurance coverage, or give medical advice. For any of those, say you'll pass it to the team and offer to take their email.
Boundaries are also where tone is most visible, because a refusal is the most emotionally loaded thing the assistant says all day. A curt refusal reads as stonewalling; a hedged one reads as evasive. Write the sentence yourself.
Two boundaries are worth setting even if nothing else is: competitors, and anything a regulator would care about in your industry.
4. Set the length and format once, for both channels
Chat replies and spoken replies read from the same Instructions field — convenient, and also the most common source of an assistant that reads well and speaks badly.
If voice is on, or might be later, avoid formatting that only makes sense on screen. Bulleted lists, bold headings and tables all become an unlistenable monotone read aloud. "Answer in short spoken sentences; use a list only when the visitor asks for one" covers both surfaces. More on that in Choosing a Voice for Your Assistant.
5. Keep facts out of the prompt
The strongest instruction you can write is one that will still be true in six months.
Opening hours, prices, policies, product specs and the shipping table do not belong in Instructions. They belong in the assistant's knowledge base, where you can replace them without touching the persona. A prompt that has quietly drifted out of date is worse than no prompt: the assistant states the stale figure confidently, in your voice.
The rule of thumb: behaviour goes in the prompt, facts go in documents. Building Your Knowledge Base covers the other half.
6. Use the AI draft, then cut it
The AI button in the AI Behavior section header opens a small panel. Describe the assistant in a line — "friendly support assistant for a bookkeeping firm" — and click Generate. Your assistant's name is sent along as context.
Two things to know before you press it:
- The generated text replaces whatever is currently in Instructions. Copy anything you want to keep first.
- Generation is rate-limited per account per hour, so it is not a slot machine.
The output is a scaffold with your specifics missing. The useful move is to generate once, then spend five minutes deleting the parts that are not true of your business and adding the two or three rules that are.
7. Set the tone of the very first screen
Persona is not only what the assistant says once it is talking. Two settings decide what visitors expect before they type a word.
Suggested Questions (on General, under Chat) are clickable prompts shown after the welcome message — up to five, 200 characters each. Use the questions people actually ask in the first ten seconds; they communicate scope far better than a welcome line. "Do you take same-day appointments?" tells a visitor what this thing is for. "How can I help?" does not.
The Welcome Message on the Appearance tab is the caption beside the orb, not the assistant's opening chat message — the opening chat message comes from the greetings on General. It is a common mix-up, and it is covered properly in Make the Widget Match Your Brand.
If you serve more than one language, Greetings by Language lets you set a per-language welcome, matched against the visitor's browser locale. Only the greeting is swapped; the conversation still follows your Response Language Mode.
8. Leave temperature alone until you have a reason
Temperature is a slider from Precise (0) to Creative (1). Low values give consistent, predictable answers; high values give varied ones. Support, booking and front-desk assistants want the low end. Move it only if answers feel robotically identical or unpredictably off-script, and move it one notch at a time. It is not a tone control — it will not make a badly briefed assistant sound warmer, just inconsistent.
Testing tone, honestly
Reading your own prompt back tells you nothing. Ask the assistant five things and read the replies as a customer would:
- A question it should answer easily.
- A question it should refuse.
- A frustrated, badly punctuated message.
- Something adjacent but out of scope.
- A question whose answer you know is not in its knowledge yet.
Then change one line and run the same five again. Once it is live, the better source is the real transcripts — the conversations you disliked tell you exactly which line is missing. Reading Your Assistant Analytics covers where to find them.
The loop that works is small and boring: read a bad answer, add one line that would have prevented it, save, move on. That beats writing 2,000 characters up front, every time.
You can try any of this on a live assistant at hiroi.ai — changes take effect on the next page load, so there is nothing to redeploy.