The overlay
The overlay is a small window that floats above everything else, for use while you are on the call. It shows a cue — a short head and a few beats — and never a paragraph.
When cues do not appear
A cue that does not fit the shape is thrown away rather than trimmed — half a
cue on a glance surface is worse than none. whispr doctor reports how many
were refused and the commonest reason, which is the part worth acting on:
- prose — the model is explaining itself rather than writing a cue. Give the
anchorrole a different model. - beats-too-long — the cues are good and do not fit the window. The same model with a shorter prompt usually will.
A refusal rate above zero is ordinary. None showing at all is the warning.
Why it is not the answer
There are two or three seconds between the other person finishing and you being expected to speak. That is not enough to read a paragraph. And reading generated prose aloud sounds exactly like reading generated prose aloud, which is worse than saying something ordinary in your own words.
So the overlay's job is to remind you, and the main window's job is to answer. The full answers are not missing — they are in the Live view and in History, which is where you should be looking if you are reading rather than glancing.
For the same reason there is nothing to click on a cue. Choosing an answer lives in the main window, because a card with buttons on it is a card that invites reading.
What is on it
The question, dimmed and on one line — enough to confirm which question this is about, not enough to invite reading it again.
Then the cue: a head in capitals, and up to five short beats separated by dots. A cue tagged prep came straight from your own material rather than from a model, and arrives first because it needs no model call.
Along the bottom, one dot per answer engine, showing whether it is idle, streaming, done or failed. A glance tells you whether the main window has something in it yet.
Showing and hiding it
⌘⇧K (Ctrl+Shift+K on Windows) toggles the overlay from anywhere, including
while the video call has focus — unless another app already holds that
combination, in which case the operating system
gives it to them and Whispr does not get it. The Overlay button in the title
bar always works. Drag the window by the strip along its top edge.
Saying a cue was good
⌘⇧L (Ctrl+Shift+L on Windows), while a cue is on the overlay, votes for it.
The strip confirms for a second or two and then goes back to the cue.
That vote is the only thing that can move Whispr onto a better cue engine. Two
engines write every cue and the fastest one wins the race — but fastest is not
best, and nothing except your vote knows the difference. Until some cues are
rated, the recommendation in Settings will keep saying there is not enough to go
on, however many meetings you run. whispr cues prints the table and the
verdict, including how many votes there have been.
It votes for the cue you were looking at, not the newest one. If another app already owns the key combination, the overlay prints no key at all — the vote is still there as a button on the answer card in the main window.
It stays out of your screen share
The overlay is marked as protected content, so screen-sharing and recording tools capture the desktop without it. This is the main reason it is a separate window rather than a panel in the main one.
When nothing appears
A cue is only shown when it comes back in the required shape. A cue engine that
answers with a paragraph, or with an explanation of itself, is refused and
nothing is rendered — nothing is better than prose on a surface you are glancing
at. You will see the question and a … where the cue would be.
If that is happening often, whispr doctor reports the rate and the commonest
reason across the install. Settings → the cue-engine table, and whispr cues,
break it down per engine — which is the grain that decides what to do, because
two engines can refuse the same number of cues for opposite reasons:
- wrote prose instead of a cue — this model cannot hold the format. Give the
anchorrole a different one. - wrote beats too long for the window — the cues are good and do not fit. The same model with a shorter prompt usually will.
Nothing appears during practice either. That mode has its own screen, and the overlay says so.
The hint line
One dim line along the bottom says where the rest of the app is. That is the whole of the in-call help, deliberately: the only question anyone has during a call is where the rest of it went.
The ? in the drag strip hands you back to the main window with this page open.
It does not expand the overlay — anyone who needs to read documentation
mid-interview has already lost the two seconds.