Use AI to prepare for an interview: 3 prompts and a mock run shown

Can AI help me prepare for job interviews?

Yes, and the three uses that hold up are predicting the questions from the job ad, a mock interview with the model playing the interviewer, and a critique of each answer against the STAR shape. In Greenhouse's 2025 Workforce & Hiring Report (2,200 job seekers in the US, UK and Ireland, published 15 July 2025) 45% said they use AI to prepare for interviews, so the person across the table assumes you did. Fewer do it well: in Kickresume's 2025 product data, about 2% of 1.22 million users of its AI features touched interview prep, against 64% who ran an ATS check. Start with the ad. Paste the whole posting, never just the title, and send this:

Here is a job ad. Read the requirements, the "who will love this job" list and the pay band. Give me the 10 questions this hiring manager is most likely to ask, ordered from most to least likely, each tied to the line in the ad that makes it likely. Mark the three that probe the weakest match between the ad and my background, which is: [two sentences about you]. Do not write answers.

I ran the interviewer version of that prompt on 27 September 2026 against a live posting, Greenhouse's own Program Manager, Product & GTM Enablement role (remote US, $88,100 to $101,960, posted 15 September 2026), on Claude Sonnet 4.5, telling it my background was three years in customer success and no formal instructional design. Its first three questions: a behavioral one, "Tell me about a time when you had to train a team on a complex product change or new feature", a situational one about Product changing a core capability two days before a rollout you built the curriculum for, and the gap question: "How would you approach building a structured curriculum for, say, a new AI-powered feature that sales reps need to demo, given you haven't formally done instructional design before?" That third question is the one to rehearse hardest. The model found it in the ad's "You should have" list in one pass, and a human interviewer will too.

Page one also agrees on a step the three prompts skip: research the company and prepare the questions you will ask. Anthropic's candidate guidance names both ("Use Claude to research Anthropic, practice your answers, and prepare questions for us") and Grow with Google's tips page puts company research first. With web search on, send: "Using only this company's own site, its press releases and its most recent quarterly update, list the five facts about its products, customers and last year that a candidate for this role should know, with a link for each. Then draft three questions I could ask the interviewer that a stranger to the company could not." Open every link before you rehearse; any date, customer name or product you plan to say out loud gets checked against the company's own page first.

How do you run a mock interview with AI as the interviewer?

You run it in a fresh chat with the model told to ask one question at a time, wait, and never answer for you, and you answer out loud, in voice mode where your plan has it. The interviewer prompt:

You are the hiring manager for the role in this ad: [paste the ad]. Interview me for 30 minutes. Ask one question at a time and wait for my answer. Mix behavioral questions tied to specific lines in the ad, one situational question about something going wrong, and a follow-up whenever my answer has no number or no result in it. Be polite and a little skeptical. Do not answer for me, and give no advice or scores until I type "debrief". Start with "Tell me about yourself".

Voice is where the practice gets honest, because typing hides the pauses and the "um". As of September 2026, ChatGPT's Live voice gives Plus subscribers 3 hours per rolling 24 hours on GPT-Live-1 and free accounts "Limited access to GPT-Live-1 mini". Google's Gemini Live runs in the Gemini app on Android and iPhone but "isn't available in the Gemini web app", and Google's own interview tips page says to use it to "simulate a real-time, conversational interview". Claude's voice mode is a beta on every plan including Free, on mobile, desktop and web. In Whizi, voice mode is on every paid plan and metered in minutes, 10 a month on Starter, 80 on Pro and 500 on Powerhouse, with a 15 minute cap per session, so a 30 minute mock is two sessions and needs Pro or above (how voice mode works).

Two rules make the mock worth the minutes. Answer without notes, because the point is finding the story you cannot yet tell in 90 seconds. And stop after four to six questions, then type "debrief", because the transcript of those answers is what you grade next. Recruiters run the mirror image: AI for recruiters has the phone screen prompt they use, with one question per must-have. The resume prompt pack covers the steps before the interview, from tailoring the resume to the cover letter.

How do you get useful feedback on your interview answers from AI?

Have a second model grade the transcript, and grade it against the STAR shape rather than asking for a better answer. The model that played the interviewer has been agreeing with you for twenty minutes; a fresh one has no stake in the conversation. Copy the transcript into a new chat, on a different model if you have two, and send:

Here is a transcript of a mock interview. For each of my answers, in order: name the question, score Situation, Task, Action and Result out of 5 each, quote the sentence where I stated a measurable result or write "no result stated", and give the one change that would raise the lowest score. Then list the three answers I should rehearse again. Do not rewrite my answers.

MIT's career office splits a STAR answer 20% situation, 10% task, 60% action and 10% result, with the note that "most of your response should focus on your Actions". Most first-pass transcripts run the other way: a minute of context, a sentence of action, no number at the end. The grader's "no result stated" entries are the fastest fix there is, because each one is a story you go and find a number for.

In a chat app with several models, the grader can be a different model in the same window (switch models mid conversation). On a single-model subscription, open a new chat so the grader never sees the interviewer's context. The "do not rewrite my answers" line is the important one, and the next section is why.

Can interviewers tell if you are using AI?

Yes, when what they hear is a memorised script. In Greenhouse's November 2025 survey of 1,236 recruiters and hiring managers, 65% said they had caught applicants using AI deceptively, and the most common tell was "Reading from AI-generated scripts (32%)". Memorising model-written answers is the one use of AI prep that backfires, and it backfires for a mechanical reason: a model writes complete paragraphs and interviewers ask follow-ups. The second question lands off the script, and the pause gives it away.

Employers have started writing the line down. Anthropic's guidance to its own candidates, dated 10 July 2025, encourages Claude for research and practice, then says of live interviews: "no AI assistance unless we indicate otherwise". Candidates and managers disagree about where that line sits. In Greenhouse's 2025 report, only 19% of US candidates said "using AI in a live interview crosses the line", against the 65% of managers who say they have caught it.

So the division of labour is: the model predicts, interviews and grades; you write the answers, as bullet points, never as sentences. Three bullets per story, the number at the end, said differently every run. A bullet you can say five ways is preparation. A paragraph you can say one way is a script, and the interviewer's second question is built to leave it.

The same rule holds one step earlier and one step later. AI for resumes and cover letters covers the tailoring that gets you the interview, and use AI to ask for a raise covers the conversation after you have the job, with the same second-model trick playing the manager.

Are dedicated AI interview prep tools worth paying for?

A dedicated tool is worth paying for when you need the camera: feedback on eye contact and body language is the one thing a chat subscription cannot give. For the three uses above, the chat model you already pay for does the same work. Prices are from each vendor's own page, in US dollars, read 27 September 2026.

Two rows need a note. Final Round AI's headline feature is the copilot that feeds you answers during a real interview, which is the use Anthropic's guidance bans and the one 32% of hiring managers say they have caught. And Google's Interview Warmup, the free tool most 2024 and 2025 guides recommend, is gone: on 27 September 2026 its address redirected to a general tips page dated 11 December 2025 that recommends Gemini Live instead.

Whizi, which publishes this page, loses here on two counts: no video mock and no eye-contact feedback, so a candidate whose last rejection mentioned presence should take Big Interview's 30 days for $39 over any chat plan, including ours. What a chat plan buys instead is the two-model step, since every paid plan puts Claude and GPT in one window. The decision rule: pay for the camera if the feedback you keep getting is about delivery, and spend the money on voice minutes if it is about examples. Either way, the thing that moves an interview is the gap question from the run above, answered in 90 seconds without notes. Practice that one until the pause is gone.