Prompts
Give the assistant a voice, a system prompt and clear refusal behaviour — then aim a profile at the visitors it is for, roll it out gradually, and see how it did.
This is where you shape how the assistant sounds and behaves. Each part — its identity, its guided behaviour, what it won’t discuss, the language it replies in — is a card showing what it is currently set to; open a card to change it, then save the profile once.
Persona & voice
Give the assistant a name and a voice tone — warm, formal, playful, or your own custom tone — and add brand phrases it should favour.
Advanced instructions
Free-text instructions for anything the guided settings can’t express. Optional — most workspaces never need it. It composes beneath the platform safety layer, so it can shape how the assistant reasons and responds but can never relax that layer.
Fallbacks & restricted topics
Choose what the assistant does when it can’t help — apologise, ask a clarifying question, or your own behaviour. Then pick the topics it should decline from a ready-made list, add your own wording for anything the list doesn’t cover, and say how it should decline. Picked categories are worded for you in whichever language the visitor writes in; your own wording is used exactly as you typed it.
Reply language
One setting decides the language the assistant answers in. Follow the visitor and it replies in whatever language they write in. List the languages you serve — in order — and it replies in the visitor’s language when it is on your list, or in your first choice when it is not, so nobody gets an answer they cannot read. Or pin a single language and every reply uses it.
Profiles
A workspace can hold several named profiles, but a conversation always gets exactly one — they are never blended. One profile is your live profile: what every visitor gets unless another one is aimed at them. Give each a short description so the next person knows why it exists, and switch a profile off to park it without deleting it.
- Starting from a template gives you more than a voice: it also declines legal, medical and financial advice from the first save, picks how the assistant says no, and sets what it does when it can’t help. Each template says what it will set before you choose it, and everything stays editable afterwards.
- Every save keeps a version. Open the history to see exactly what changed, and restore an earlier version in one click — restoring adds a new version rather than erasing anything.
- Compare any two profiles side by side to see which settings differ, with their two composed prompts next to each other.
- Export a profile to a file and import it into another workspace. The file carries the wording only — never who it is aimed at, and never which profile is live.
- The list shows each profile’s last 7 days at a glance — how many chats it served and how they felt — so you can see which one needs attention without opening it.
Who gets this prompt
By default a profile is for everyone. You can instead aim it at named audiences, at specific signed-in people, at conversations about particular capabilities or services, or at visitors who are not signed in. When more than one profile could apply, the most specific rule wins: a named person beats an audience, an audience beats a capability rule, and any targeted profile beats the untargeted default. If two are equally specific, the priority you set breaks the tie.
Not sure who a profile would reach? Open the simulator, describe a visitor, and it shows which profile wins and why — including every profile it considered and the reason each one lost.
A page can also ask for a specific profile by name from its embed snippet, using the data-prompt-profile attribute — but only profiles you have published for that purpose can be named, and anything else falls back to your live profile.
Gradual rollout
Once a profile is aimed at someone, you can hand it to a share of them instead of all at once. Set a percentage and the rest keep whatever they would have had. The split is sticky: the same visitor stays on the same side of it for as long as the percentage is unchanged, so nobody sees the assistant change voice mid-conversation — and raising the percentage adds people rather than reshuffling everyone.
How a profile is doing
Each profile reports what actually happened in the conversations it served: how many there were, how they felt, how they ended, thumbs up and down, the questions it could not answer, and the average length. Choose a 7-, 30- or 90-day window; if your retention setting is shorter than the window you picked, the page says so and reports the period it really has.