Persona
Shape how the agent talks, what it always knows, and what it says when it can't answer.
Agent > Persona is where agent behavior is shaped: custom instructions, persistent knowledge, and the static fallback message. Everything here is Liquid-templated, so one agent can behave differently per user based on their attributes.
The agent refers to your product by your app's name, which you set in Settings > App. What language it replies in is set on the Localization page.
Agent configuration
Two Liquid-templated fields, each with a live character counter:
- Custom instructions (5,000 characters): steers tone and focus. It can't override how the agent formats responses or which elements it may reference.
- Persistent knowledge (5,000 characters): facts the agent should always know, on every page, for every user. For anything longer than a few sentences, use Sources instead.
Type {{ in either field and a list of your attributes opens, filtered by whatever you type next; pick one and the whole reference is written for you. Insert attribute above the field does the same for anyone who would rather not type it. An
attribute shows in the field as a single highlighted chip rather than raw braces, and one
press of Backspace or Delete removes the whole reference, so you cannot leave half of one
behind. A Liquid warning appears if you reference an unknown or archived attribute, but a
warning never blocks saving.
Static fallback message
A toggle, off by default. When on, the agent shows a message you write instead of guessing when it genuinely can't answer from your product's documentation. Turning it on with a blank message blocks saving until you either write one or turn it back off. The message supports bold, italic, links, and lists, and is capped at 1,000 characters.
It also accepts attribute references, the same as Custom
instructions above it: write Sorry {{ user.first_name }}, I could not find that. and each
user reads their own name. Typing {{ opens the same attribute list here as in the fields above, and Insert attribute in the editor's toolbar drops a reference at your cursor. A message that is not valid template syntax is refused when you save, so a broken
reference can never reach a user.
Previewing live
Preview live, next to the page title, opens a page of your choice in a new tab with the real ARCY widget on it, trained on your Sources and wearing your active theme, even when ARCY is not installed on that page yet. ARCY fetches the page for you and adds the widget to it, so what you see is the published state of this product: the persona you last published to the active environment, its Sources, and its theme.
- Nothing is recorded. Your own conversation on a preview page never touches analytics.
- One page per preview. Links inside the page lead to the real site, where the widget is not present. Open another URL to preview another page.
- Verified domains are one click. Any other URL is reserved for ARCY's own team.
- The preview link works for thirty minutes, reloads included.
Saving and publishing
Persona follows the same Draft and Publication model as Flows: one save button for the whole draft, and a separate Publish action per environment. Publish is disabled while you have unsaved changes. Save first, then publish.
If someone else saves the same draft while you have it open, your save is refused rather than silently overwritten; reload to pick up their version before continuing.
See also
- Sources: attaching longer reference material the agent can retrieve
- Customization: the widget's visual theme, a separate saved (not published) surface