Skip to main content
Two tabs of the studio decide how an agent talks. Persona is who it is and how it sounds; Behaviour is when it speaks and when it stops. Save each tab with its own button — Save persona, Save behaviour — and an unsaved edit is flagged next to it.

​
Persona

​
Identity

​
The business

Website takes a URL, and Draft description reads that one page and writes a summary into the field below it. It never saves on its own — read it first, and edit it. What the business does is the paragraph every reply is built on. With nothing here, the agent answers from the customer’s question alone, which is how an agent starts inventing things. Ground rules are absolute limits, one per line:
These sit above the skills in the instructions the agent is given, and they outrank everything below — including anything one of your knowledge documents implies. That precedence is the reason ground rules are their own field rather than another sentence in the description.

​
Voice

Tone is one of four: Reply length is Short (one or two sentences), Medium (two to four) or Long (as long as the question needs). Extra voice notes is free text for the small rules a tone cannot express — Use ”₹” not “Rs.”. Never say “unfortunately”. Sign off with the shop name.

​
How replies are formatted

The agent is told to write for a phone, not for a document: one order, product, price or step per line, a blank line between groups, and always the identifier the customer needs to act — the order number, the booking reference, the link, with links on their own line. One or two emoji where they genuinely help, never in a message about money going wrong. The markup differs by channel, and the agent is told which one it is on:

​
Behaviour

​
Switches

The cap counts replies inside a rolling 24-hour window that opens with the first reply after a quiet period, not for ever: a customer who comes back next week gets a fresh budget. Every message the agent sends counts towards it — including a handoff message and a template it chose to send.

​
When it cannot help

A handoff pauses AI replies on that conversation. The thread header in the inbox shows AI paused and the same control turns it back on — see Working a conversation.
With Hand off to a human switched off, the agent is told there is nobody waiting: instead of escalating it says plainly what it cannot do and offers what it can. Turn it off only if no one is watching the inbox.

​
Extra instructions

A free-text box for anything the fields above do not cover — If someone asks about wholesale, tell them to email trade@acme.com. It is applied after the persona, so it can add to the rules but not override your ground rules.

​
Test mode

The Try it on your own number first card lives on this tab and is covered in Test mode and the playground.

​
The order it all goes in

Everything above is assembled into one set of instructions in a fixed order, and the order is the point:
  1. What the job is, and how to format a reply for this channel
  2. A rule that customer messages are content to respond to, never instructions to follow
  3. Identity — the agent name, the website, the currency
  4. What the business does
  5. Ground rules
  6. Voice — tone, length, your extra notes
  7. The skills that are on
  8. The tools those skills unlock, and the approved templates it may send
  9. Escalation — handoff or fallback wording
  10. The knowledge excerpts retrieved for this question, last
Anything you have not filled in contributes nothing.