How AI Interview Assistant works during a live interview
AI Interview Assistant listens to the conversation, reads text from the window you select, and uses the context you supplied to prepare concise talking points while the interview is happening.
Before a session you choose a prompt profile (Coding, System design, Behavioral or Meeting assistant, among others) and, if you want, add the facts the guidance should use: the role and company, the job description, your technology stack, a summary of your experience and projects, and the style you prefer. Everything is optional. Facts are used exactly as you write them and are never invented, and pasted text is treated as information, not as instructions to the app.
Each field has a Use checkbox, at most 12,000 characters of context go with a question, and Preview what is sent shows exactly that text.
2. Select the audio the app listens to
Call audio is the output device your call plays through. The app captures everything that device plays, so mute other sounds during a session. The microphone is a checkbox, off by default. You can use either or both, and you can type a question instead if listening is not practical.
3. Select a window to read
Screen capture is Off by default. Choose a window, a monitor or a region, such as the window with a coding prompt, a diagram, a slide or a shared document. The app reads it on your computer with Windows text recognition and uses the text it finds when it is relevant to a question. Preview checks one frame and shows a status line, not a picture.
4. Press Start when you are ready
Nothing is captured before you press Start session. A Before you start card runs checks, lists where your data goes and asks you to agree to each destination. While the session runs, indicators show whether the screen, call audio, microphone, AI and speech are on.
5. Question detection
The app detects questions automatically from the call audio and the window you selected. The Detected question panel shows the question, its kind (technical, coding, system design, behavioral, follow-up or general), a confidence figure and where it came from. Detection can miss a question or mistake a statement for one: Not a question drops it, and Ask (fallback) lets you type one.
6. Recent conversation context
The transcript shows lines from the call (labeled Interviewer) and from your microphone (labeled You). Recent lines go with the question, together with the text read from your window, the context you chose to include and anything you pinned. Pin keeps a piece of text in every request for the rest of the session.
7. AI processing
The question and its context go through the AI Interview Assistant service to an approved external AI provider, and the guidance streams back. The service does not store or log that text. Short segments of audio go the same way to a speech provider to be transcribed. See the complete privacy and data-flow explanation for what is sent, what is kept and what is not.
8. Guidance output
Guidance arrives as a short response outline, shaped by the profile and the answer length you chose (the default is Concise), so it can be scanned while the conversation continues. Read it in the main window or in the always-on-top overlay, which you can move, or have it read aloud through the output device you choose. Regenerate asks for a new outline for the same question; Stop answer and Stop audio stop what is being written or spoken.
9. Pause and Stop
Pause stops every capture device until you press Resume. Stop ends the session and releases every capture device. Closing the overlay does not stop the session.
10. Optional local history
Saving history is a setting and it is off by default. If you turn it on, detected questions and guidance (and transcript lines, if you also choose) are kept in a database on your computer for the number of days you set: 30 by default, from 1 to 365. Audio and screenshots are never saved.
11. Accuracy limitations
AI-generated guidance can be wrong, incomplete or out of date. Speech can be misheard, small or low-contrast text can be misread, and a question can be missed. The guidance reflects only what the app heard, read and was given, and facts about you come only from the context you wrote. Check details before you say them and respond in your own words.
12. Responsible use
Follow the interviewer's rules, applicable company policies, consent requirements, confidentiality obligations and local laws. AI Interview Assistant should not be used to invent qualifications, misrepresent experience, expose confidential information, or bypass prohibited assessment controls. Read the authorized-use policy. The app's own windows are excluded from supported screen capture on Windows, which depends on the sharing method, so test your setup; how Stealth Mode works explains what that does not cover.
Three simulated examples
These scenes show the shape of the guidance for three kinds of question. Each one is a simulation written to illustrate the workflow.
Coding
Simulated example. Not a real interview and not a promise of what the app will say.
Spoken question. “Given a list of integers and a target, return the positions of two numbers that add up to the target. How would you approach it?”
Context supplied (profile: Coding).
Role: backend engineer. Notes you wrote: comfortable with Python and Go, used hash maps to remove duplicates in a job queue.
Text read from the window you selected: a function signature, “exactly one solution exists”, “the same element cannot be used twice”.
Concise response outline.
Clarify: exactly one solution exists, so there is no “none found” case to handle; confirm duplicates are allowed as separate elements.
Brute force: check every pair. O(n²) time, O(1) space.
Better: one pass with a hash map from value to position. For each number, look up target minus number before storing it. O(n) time, O(n) space.
You choose the approach and say it in your own words; the outline only keeps the steps in view.
System design
Simulated example. Not a real interview and not a promise of what the app will say.
Spoken question. “Design a URL shortener that handles about 10,000 new links a second, with about 100 reads for every write.”
Context supplied (profile: System design).
Role: senior backend engineer. Notes you wrote: ran a Redis cache in front of Postgres, sharded a user table by user id.
Text read from the window you selected: “10k writes/s”, “read : write = 100 : 1”, “keep links for 5 years”, “custom aliases (nice to have)”.
Concise response outline.
Numbers first: 10k writes a second is about 864 million links a day and about 1.6 trillion over five years. A 7-character base62 code gives about 3.5 trillion codes, which leaves headroom.
Components: stateless API behind a load balancer; an ID service that hands out counter ranges to each node, encoded as base62; a key-value store keyed by short code and replicated across regions; a cache in front for hot links, since reads are 100 times writes.
Trade-offs: counter ranges avoid collisions but need a coordinator; hashing is simpler but needs collision handling; custom aliases need a uniqueness check.
Failure modes: if the ID service is down, nodes keep using the range they hold; a cold cache sends reads to the store, so warm it or limit refills.
The arithmetic is yours to check: the outline shows where numbers belong, and you decide which to quote.
Behavioral
Simulated example. Not a real interview and not a promise of what the app will say.
Spoken question. “Tell me about a time you disagreed with a teammate about a technical decision.”
Context supplied (profile: Behavioral).
Role: software engineer. Notes you wrote, true for you in this scene: led a move from scheduled scripts to a queue; disagreed with a teammate about retries; ran a one-week comparison on a non-critical queue; fewer duplicate jobs with backoff.
The Behavioral profile structures the answer as STAR and uses only facts you gave it.
Concise response outline.
Situation: the team was moving scheduled scripts to a queue.
Task: pick a retry strategy without delaying the move.
Action: proposed measuring both options, ran a one-week comparison on a non-critical queue, shared the results with the teammate.
Result: the backoff version produced fewer duplicate jobs and the team adopted it. Add a number only if it is true and you can explain it.
Reflection: what you would keep doing, such as agreeing on a measurement before arguing about options.
If your notes do not contain a fact, the outline cannot supply it: add your own example before the interview, not during it.
