Persona and Safety controls how the assistant behaves: its tone, its language, and the limits it works inside.
Tone
A short description of how the assistant should sound. The default is professional, efficient and slightly witty.
Describe it in a few words rather than a paragraph. "Direct and warm" produces a consistent voice. Four sentences of nuance produce an assistant that is somewhat all of them and reliably none.
Match how your organisation actually writes, not how you would like to. If your team writes short and plain, an assistant that writes elaborately will have every draft rewritten before it is sent, and the AI will quietly stop being used for drafting.
Response style
How much it says. Concise answers a question; detailed explains around it.
Concise is right for an assistant somebody uses often. Detail is welcome the first three times and tiresome by the tenth, and people stop reading past the first line long before they stop asking.
Language
Ten languages are available, including English, Arabic, French, Spanish, German, Portuguese, Hindi, Chinese, Japanese and Korean.
This sets the default. It matters most where the assistant drafts anything a customer reads, and it is worth setting deliberately in any organisation where the working language and the customer language differ.
Custom instructions
Free text that applies to everything the assistant does. The most powerful control on the page, and the easiest to overload.
Good instructions are rules with consequences. "Never quote a price without saying it excludes VAT." "Always give the amount in the customer's currency first." "Refer to our service as a subscription, never as a licence."
Bad instructions are aspirations. "Be helpful." "Be accurate." "Think carefully." These consume space and change nothing.
Keep it short. Five sharp rules are followed. Thirty are averaged.
Content policy
Three levels, and the names are honest about what they trade.
Strict. Formal, no humour, no speculation. Right for regulated work, and for anything the assistant drafts that goes to a customer unedited.
Moderate. Professional with some flexibility. The default, and correct for most organisations.
Permissive. Creative, informal, experimental. Right for internal brainstorming and marketing exploration, and wrong anywhere near a customer.
Strict is not the safe default. It is the right setting for a specific kind of work. Applied to an internal assistant helping people think, it produces answers that refuse to engage with hypotheticals, which is most of what anybody wants to ask about.
Blocked topics
A list the assistant will not discuss.
Useful and narrower than it appears. It is right for genuine no-go areas: legal advice, medical opinion, anything your organisation has a policy against commenting on.
It is not a security control. Blocking "salaries" as a topic does not stop somebody who can see salary data from seeing it elsewhere, and it does not stop somebody who cannot see it from being correctly refused. Access is governed by permissions, and it is governed the same way whether somebody asks the assistant or opens the page.
Use blocked topics for things you do not want discussed, and permissions for things you do not want seen. Confusing the two produces a system that feels controlled and is not.
Persona applies further than you think
The settings on this page are not only for the chat window.
They shape anything the AI writes: a drafted reply in the inbox, a summary after a meeting, the chase an agent produces at six in the morning, an answer given to a customer over WhatsApp.
That reach is the argument for setting it deliberately rather than leaving the default. An assistant that sounds slightly off in a chat window is a curiosity. The same voice in a message a customer reads is your organisation's voice, whether you chose it or not.
It is also the argument for testing changes against drafted output rather than only against questions. Tone that reads as friendly in an answer can read as flippant in a payment chase, and the payment chase is the one that matters.
Getting it right
Change one thing and test it. The whole page is one prompt in effect, and adjusting tone and instructions together makes the result unattributable.
Test on real questions. Persona changes read well in the abstract and produce surprises on actual work, because the questions people ask are stranger than the ones you invent.
Ask somebody else to read the output. You wrote the instructions, so you read the intent into the answer. Somebody who did not read what is actually there.
Multilingual reality
If you operate in more than one language, the setting is a default rather than a constraint.
Somebody writing to the assistant in Arabic gets an answer in Arabic. The setting decides what happens when there is no signal: an unprompted summary, a drafted email, an agent's output.
Set it to the language your customers read, not the one your team speaks internally, if the AI drafts anything outbound. Internal answers follow whoever asked.
Defaults worth keeping
Not everything here needs changing.
The default tone is deliberately unremarkable, which is right for most organisations. Moderate content policy suits nearly everybody. An empty blocked-topics list is correct unless you have a specific reason.
The one to always set is custom instructions, because that is where your organisation's actual rules live, and no default can guess them.
What this page does not do
It does not make answers accurate. Tone and truth are unrelated, and a confident wrong answer in a lovely voice is worse than a clumsy right one. Accuracy is Teaching the assistant.
It does not control what data the AI can reach. That is permissions, and it applies identically whether a person asks through the assistant or through the interface.
It does not apply to agents' judgement. Agents inherit the persona for anything they write, but what they look at and what they do is set on each agent.
Did this answer your question?
No, ask a person