Sobes Copilot setup before an interview: a 10-minute checklist

Ten minutes until the call. The meeting link is open, you've reread your resume, Copilot is running. The most frustrating failure at this point looks like this: the interviewer is already asking the first question, and the transcript is empty because the app can't hear system audio. You won't fix that on the fly while introducing yourself.

A technical interview in the leaked interviews from our bank averages 18 to 20 questions, and the first one arrives in the second minute. So everything below is done before the call, in order, and takes about ten minutes. If you're setting Copilot up for the first time, go through the quick start in the docs first and come back to this list right before the meeting.

1. Check the sound: the interviewer and your mic

The audio recording settings have a "Check sound" button. The app plays a test question through your speakers, as if an interviewer asked it, and shows two rows: the interviewer's voice (system audio, meaning whatever plays in your speakers or headphones) and your microphone. Each row ends up as heard or not heard. Say a few words out loud while the check is running so the mic gets tested too.

If everything works, you'll see the text Sobes heard and a sample hint below it. That means the whole chain is fine: audio captured, transcribed, answer delivered.

Keep the audio source on system audio plus microphone. The mic matters for more than transcribing your side: the post-interview summary scores your answers specifically, and without a mic recording it comes back with no scores.

2. On macOS: screen recording

On a Mac, system audio depends on the screen recording permission. Without it Copilot hears only the microphone, and screenshots show nothing but the desktop. You grant it in System Settings, under Privacy & Security, Screen & System Audio Recording.

There's one catch people trip over most often: macOS applies the permission only after the app restarts. Turn the toggle on, press "Restart", and run the sound check again.

3. Activate the context for this role

Only one context can be active, and it applies to every answer in the meeting. If yesterday was an HR screening and today is a Go technical round, make sure the right one is on. The active context's name is shown in the overlay toolbar, and you can switch it right there.

Before an important meeting, open the context and use the test chat: ask a question as the interviewer would and look at the tone and depth of the answer. We covered the context fields in What contexts are and how they improve AI answers.

The job description isn't set in the context. You add it when creating the meeting, as a link or as pasted text, and the assistant then takes the role's requirements into account.

4. Create the meeting from a tracker card

If the process already lives in the application tracker, pick it under "Tracker process" when you create the meeting. The vacancy comes from the card, and the meeting with its summary gets attached to the process. Three stages later you'll open the card and see every conversation with that company in one place.

There's a practical audio benefit too: the company name and key skills from the meeting's vacancy are added to the speech recognition hints (for Deepgram, Soniox, Inworld, and Grok STT). The name of a framework or the company's product is less likely to be transcribed wrong.

If the stage on the card has a date and time, the app will offer to start a meeting for that card five minutes before it begins.

5. Set the languages

Languages are chosen in the meeting, not in the context: in the overlay it's the flag button next to the context picker. The menu has four settings:

• Answer language. The assistant answers in it no matter what language the interviewer speaks. "Auto" means the language of the interviewer's last phrase.
• Answer translation. Up to two languages, and the translation appears under the answer.
• Transcription language. Russian by default. If the interview is in English and the transcription language is still Russian, recognition gets noticeably worse.
• Transcript translation. It translates the interviewer's phrases and leaves yours alone.

A common setup for an interview in English when it isn't your first language: transcription in English, answers in English, answer translation into your native language so you can grasp the meaning before you speak. The format itself is covered in Interviewing in English.

6. Recording mode and transcription model

VAD mode fits most interviews: you turn recording on once (Ctrl/Cmd + R), and from then on the app finds speech by itself and sends phrases for transcription. If you'd rather decide what gets recorded, there are Start / Stop and One-Shot.

Speech is recognized by Parakeet by default. If a model doesn't respond, the server tries the next one, so the phrase isn't lost. With Deepgram, Soniox, and Inworld models you can turn on realtime recognition: text appears as people speak, and the answer arrives sooner. Models are compared by latency and accuracy in How to choose a speech transcription model.

Ten minutes before a call is a bad time to switch models. Check the one you've already used.

7. Notes and documentation for this meeting

When you create a meeting, the notes button with the folder icon limits search to the folders you pick. That helps if you keep several versions of your stories for different roles. Installed documentation is selected in the same place.

You can check in advance whether the right note will be found: the notes section has "What Copilot finds". Type a question the way an interviewer would ask it and see which fragments would go into the hint. More on notes and pet project code: How to teach Copilot to answer about your experience.

8. Check invisibility while sharing

Open the compatibility check page in the same browser you'll use for the call, start sharing, and choose the entire screen rather than a window or a tab. If Copilot isn't visible in the preview, the interviewer won't see it either. If it is visible, try another monitor or share a single window.

Use a Chromium-based browser for calls: Chrome, Edge, Brave, Yandex Browser. The Zoom desktop client needs "Advanced capture with windows filtering" enabled in its settings first, while the web version needs no setup. The list of services we've tested is in the docs.

9. Four hotkeys you need in the first minutes

Press each one once before the call. If another app intercepts a shortcut, you'll find out now and not in the middle of live coding.

If the interviewer's audio isn't recognized

If the call is already running and there's no transcript, don't go silent and don't fiddle with settings in front of the interviewer. Take a screenshot of the task and send it (Ctrl/Cmd + H, then Ctrl/Cmd + Enter) or type the question briefly into the input field. That's enough to get you to the next pause.

A dry run the day before

The most reliable check takes five minutes and is done a day before the interview. Create a test meeting, turn recording on, and play any video in the browser where someone asks questions. The transcript should show up in the chat, and the questions should get answers. Then open the compatibility page and make sure the overlay doesn't show up in the share.

How to behave during the conversation itself, where to put the window, and what to do with pauses is covered in How not to give yourself away during interviews.