Start
The conversation in the middle, and beside it whatever waits for your yes.
Two things want your attention at once: what you are about to do, and what has been left lying in the office. The start screen separates them in space. You write in the middle, what waits for you stands to the right, and neither crowds out the other.
The layout
Two columns from 1280px, one below that. The middle is capped at reading width; the work column on the right is 420px wide and hangs off the panel on a hairline.
As long as no conversation is running, the middle reads from top to bottom: the face, the greeting (Good morning, Good afternoon or Good evening, with your first name), and below it the input. After more than a week away it says Welcome back.
The work column has four tabs, named after the screens they point at:
| Tab | What is there |
|---|---|
| For approval | Drafts waiting for your yes. Only appears while something is open, and is then the open tab |
| Viewings | Your week, compact |
| Inbox | What has newly come in |
| Tasks | Six rows at most, then the way to all of them |
Once a conversation is running, the header and the column drop away. Only the thread and the input remain.
What you can do here
Above the input sit five chips under the line I can help with:. A click fills the input with a sentence; it does not send. As a provider those are New property, Valuation, Task, Edit listing and Change status. The same five are available mid-conversation as slash commands.
On the input itself: Send, Stop, Attach image or PDF and Dictate. Enter sends, Shift and Enter make a new line. Attachments may be images and PDFs, up to 4MB each and three at most.
On the edge of the shell sit New conversation and the Recent conversations menu, grouped by Today, Yesterday, Last 7 days, Last 30 days. Every row there is a real link, so it opens in a new tab and can be bookmarked. Deleting asks once and is final.
The column can be put away with Collapse column; the choice survives a page change.
What the agent does here
Three things, visible in this order.
It says what it is doing. One line per tool: "Drafting the reply", then "Reply drafted". While something runs, three dots sit beside it. A question carries a question mark; a failure carries a warning mark and the reason in plain words below it.
It asks with a card, not a sentence. If something is to leave the house, a confirmation card appears: the wording that would go out, the recipient, the time. One click is enough. The no sits on the left so the Enter key never confirms by accident, and a click in the first fraction of a second does not count.
It puts drafts in the column. Under For approval each entry shows the subject, the kind and the whole reason. The Approve button only appears once the full text is visible.
Never approve blind
That is not an accident of the implementation, it is the rule: while the text is cut off, there is no approve button. What you approve, you have read.
Keyboard
In the approvals list: J and K or the arrow keys move the selection, Enter opens. The keys only work while focus is inside the list.
Worth knowing
Which conversation is open is part of the address (/?c=...). A running conversation does not carry it yet: the id is issued by the server, and setting it afterwards would rebuild the screen mid-typing. Reload during that and you start at the front, with the conversation under Recent conversations.
Half-written sentences stay in the input, separately per side: a provider's half sentence does not belong in a searcher's input.
Empty blocks in the column do not disappear. They stay with their heading and one sentence, so the screen has the same shape empty as it does full.