Fragments are reassembled
“So the thing I wanted to ask” — pause — “is how you handled the complaint” is one question in two pieces. The pieces are glued back into a single turn before anything looks at them.
Question detection
Knowing you were asked something is harder than answering it.
An interview is forty minutes of talking, of which maybe fifteen turns are questions. Answering the other three hundred fills your screen with noise in the one moment you cannot afford to read any. This is how the fifteen are found.
Listening and classifying cost nothing. You are charged for an answer.
Step one
Wait for the person to finish
Spoken questions do not arrive as sentences. They arrive as pieces, with pauses that look like endings.
A pause is not an ending
People pause mid-sentence to think, and an interviewer pauses more than most because they are reading your CV while they talk. Treating the first silence as the end produces an answer to half a question.
Step two
Decide whether that was a question
Locally, in a couple of milliseconds, for nothing — because this runs on every turn and most turns are not questions.
How it opens
“How”, “why”, “walk me through”, “suppose”, “imagine”, “tell me about”, “talk me through”. Interviewers phrase questions as instructions constantly, and an instruction is a question.
How long it is
Under three words it is almost never a question. “Sure”, “mm-hm”, “okay great” are the texture of a conversation, not asks.
What it is not
A list of acknowledgements and filler that must never trigger anything: agreement, scheduling, the interviewer narrating their own screen.
What it is about
Technical, behavioural or situational, which decides how the answer is shaped — a behavioural question wants a story with a result, not an explanation.
A threshold, not a guess
A turn has to clear a score before it reaches anything paid for. Everything below it is ignored silently.
Free by design
No model call, no credit, no round trip. Which is what lets it run on every single turn instead of on the ones that look promising.
Step three
Count how many questions that was
A separate decision from the one above, using different cues — and merging the two would break both.
The classifier asks whether this is a question at all, and by that measure “suppose a pod keeps restarting” counts. The splitter asks how many questions this is, and by that measure the same sentence is scene-setting for the ask that follows it. Two questions, two sets of cues, kept apart on purpose.
Two asks, two answers
“Tell me about the restructure, and also how you would do it differently” becomes two parts, shown separately and answered in order.
Scene-setting is not an ask
“Imagine a third of the team hands in notice in one month — how would you find out why?” is one question with a premise, not two questions.
The parts stay visible
You can see what it thought it was asked. If it split something it should not have, you know immediately, rather than wondering why the answer is odd.
When you want to drive
Three ways to ask it something yourself
Detection is for the call. These are for everything else.
Type it
For an interview in a room, a question that arrived over chat, or one it missed. Same answer, same context, same speed.
Paste it
A take-home prompt, a long scenario, a question from an email. Faster than reading it to your laptop.
Send your screen
One keystroke in the Windows app sends what is on screen — a coding problem, a diagram, a slide — and the answer comes back with the code. The Windows app →
Questions
About detection
How does it know something was a question?
A local pass over every turn, before anything is sent anywhere. It scores the turn on how it opens, how long it is, and whether it ends on a question mark — because spoken questions very often do not. “Walk me through your last big project” has no question mark and is unambiguously a question; “right, got it” has three words and is not.
Does it answer every single thing the interviewer says?
No, and that is most of the work. Acknowledgements, thinking noises, scheduling chat and the interviewer agreeing with you all score below the threshold and are ignored. A tool that answers everything fills your screen with rubbish at exactly the moment you cannot afford to be reading rubbish.
What if it answers half a question?
It waits for the end of the turn before deciding anything. Speech arrives in fragments — “so what I wanted to ask”, pause, “is how you handled the complaint” — and those fragments are stitched back into one turn first. Answering a fragment is worse than answering late.
What happens when two questions are asked at once?
They are split and answered in order, as separate parts you can see. “Tell me about the restructure, and how you would handle it now” answered as one blob reliably loses one half of it, usually the half the interviewer cared about.
Can I ask it something myself?
Yes. Type or paste a question at any time, which is what you want for an interview in a room, a question that came over chat, or a case where it missed something. In the Windows app a keystroke also sends whatever is on your screen, for a coding problem or a diagram.
Does detection cost credits?
No. Listening and classifying are free however long the call runs; you are charged for an answer. That is deliberate: a classifier you pay for is a classifier you would rather have answer less often, which is the wrong incentive to build into the thing.
Read next
Related pages
Watch it ignore the small talk.
Run a session, say a few things that are not questions, and see what it does and does not react to.
Create your account