Paste the lines. See which ones were questions for you.

Try it

What the first stage looks for

The rules are simple on purpose; they run on every line in a few milliseconds, and their job is to be generous — to pass anything that might be a question and drop only what certainly is not.

• A question mark, or a line that starts like a question: what, why, how, can you, could you, do you, have you.
• An instruction: tell me, describe, explain, walk me through, give me. Instructions are questions without the mark.
• An invitation to speak: I’d like to hear, your thoughts on, let’s talk about, over to you — asking without a question.
• Your name, if you gave it. A line with your name in it is probably for you, even when it is not grammatically a question.
• “You”, in a line long enough to mean something. The weakest signal; it goes to the model but would not by itself raise a card.
• Dropped: acknowledgements — mm-hm, right, okay, got it, that makes sense — and lines too short to carry a question.
• Dropped: your own lines, and everything else that is simply a statement.

What the first stage does not know is who a question was addressed to. In the example, “do you want to take the system design part” is a question, and it is for Priya. The first stage keeps it; the second stage — a model that looks at the last dozen lines, the names of the participants and your name — is what decides it was not for you. That stage does not run on this page, and the page says so, because it would need a server and it would cost money to run for every visitor.

Why it is built this way

If the first stage were strict, the copilot would miss questions, and a missed question in an interview is the worst failure a copilot can have. If it were absent, every line would go to the model and the window would fill with cards about acknowledgements. Generous first, precise second is the trade-off, and the count under the box shows it: on a normal interview, roughly a third of lines pass the first stage, and a smaller share come back from the second as a card.

What happens in the real extension after this

The lines that pass go, with the last dozen lines of context, to the model; the model returns either nothing or a short answer with the line it grew from; a confidence threshold drops the weak ones; and a similar question within a minute does not make a second card. All of that is on the question detection page. The part on this page is the first ten milliseconds.