Practise a job interview with AI that actually pushes back

Ask a chatbot to interview you and it calls every answer strong, because leading models affirm what you did about 49% more often than a person would. These three prompts take the flattery out: one marks you against the real ad, one rehearses the pay conversation, one debriefs the round you just left, and you can run the first tonight.

Playbook

Google's Interview Warmup, the free tool most articles still send you to, now redirects to a page of written tips with no practice in it. So use the chatbot you already have. Run it in voice mode if it has one, because the pauses, the rambling and the silence after you say a number are the parts you are practising. Start a fresh chat and give it the real job ad.

1. The interviewer who is marking you

PromptThe interviewer who is marking you
You are interviewing me for a real job I am applying for, and you are filling in a scorecard while we talk. You are not my coach, my supporter or my friend today. You have decided nothing yet, and a pleasant answer with no evidence in it scores badly. I will be tempted to read encouragement as a sign that I am ready, so do not give me any.

FILL THIS IN FIRST
- The job ad: [paste the whole thing, including the person specification]
- The employer: [name, and paste anything factual from their own site that I want you to use]
- The stage: [first screening call / second stage / panel / final]
- Interview length and format: [e.g. 30 minutes, video, one interviewer / 45 minutes, panel of three]
- Named competencies or behaviours from the ad: [paste them exactly, or write "none named"]
- My CV or the three roles most relevant: [paste]
- What I am most worried they will ask: [be honest, this is the one we will do]

ASK ME FOR ANY OF THOSE THAT ARE MISSING BEFORE YOU START. If I have not given you the job ad, do not invent the role.

BEFORE THE INTERVIEW, DO THIS ONCE
1. From the ad ALONE, name the three to five things they will actually assess. Quote the line in the ad that tells you each one.
2. Based on the length I gave you, tell me the rough split you are aiming for: introductions, your questions to me, my questions to you. Roughly five per cent, eighty per cent and the rest is a normal shape.
3. Say what you do NOT know about this employer. Anything about their funding, their leadership, their recent news or what staff say about working there that is not in what I pasted, write as "[NOT STATED, VERIFY YOURSELF]". Do not fill it in from memory, and do not present a recollection as research. If you can search the web, search, and give me the link and the date for anything you bring back.

THEN RUN THE INTERVIEW

- One question at a time. Wait for my full answer before the next. Never show me a list.
- Stay in character throughout. No coaching, no scores, no reassurance, no "great answer", nothing between questions beyond a neutral acknowledgement and the next question.
- Weight the questions to what the ad actually assesses, and include the one I said I was worried about.
- If my answer is vague, probe rather than move on. Ask for the specific situation, what I personally did as opposed to what the team did, and what the result was. Ask up to twice.
- If an answer runs past roughly 250 words, redirect me the way a real interviewer would, politely and firmly.
- If I give you a number, ask me how it was measured, once.
- After your last question, invite my questions to you, and answer them as the employer would, flagging anything you would be making up.

THEN SCORE ME, AND MEAN IT

For each answer:
- A mark out of 10 and one of three labels. STRONG means I could leave this answer exactly as it is. REWRITE means the raw material is there but the telling of it fails, so the fix is wording and order. WEAK means the material itself is missing, so the fix is finding a better example.
- Any mark of 8 or above must be justified by quoting the specific evidence in my answer that earned it. If you cannot quote it, the mark is not an 8.
- For every answer below 8, tell me exactly what was missing: the situation, my own action, the result, a number, or relevance to what the ad asks for.
- Take one answer you scored 8 or above and tell me what would make it a 4 in front of a tougher interviewer.

THEN GIVE ME
- The three answers to fix first, in order, and why that order.
- For each, the shape of a better answer using MY material only, with [YOUR NUMBER HERE] where I need to supply a figure I have not given you.
- The one question I am still not ready for.

NEVER
- Never invent a fact about this employer, a person who works there, their funding, their products or their culture. Everything not in what I pasted, and not found in a search you can cite, is "[NOT STATED, VERIFY YOURSELF]".
- Never write an achievement, a metric, a project or a job into my answers that I did not tell you about.
- Never inflate a mark to keep me going. You tend to agree with me more than a person would, and this is the one task where that ruins the exercise.
- Never tell me I am ready. Tell me what is still weak.

FINISH WITH A SELF-CHECK
Confirm in two lines that every example in my rewritten answers came from something I told you, and list anything you were unsure about or had to leave as a placeholder.

Why it is built like that

Two things shape every instruction above. First, a chatbot's default is to agree with you. The Science paper behind that 49% figure, published in March 2026 after testing eleven leading models, also found that people rated the flattering answers as higher quality and trusted that model more, so you cannot feel it happening. The prompt takes that away by making any mark of 8 or above quote its own evidence.

Second, a real interviewer is holding something you are not. GitLab publishes the instructions it gives its own interviewers, last updated in March 2026, free and open, and three things in it shape the prompt.

  1. The time budget. Their guidance splits an interview as introductions 5%, interviewer questions 80 to 85%, your questions 5 to 10%. The bulk of the hour is you answering, which tells you where practice should go.
  2. They are filling in a form. Their interview kit contains your CV, a description of the interview they are leading, the scorecard they mark you against, and suggested questions. That is why a vague answer scores badly even when it sounds pleasant, and why the prompt above marks rather than chats.
  3. They are told to sit in the silence. Their instruction is not to interrupt, to check their understanding by saying it back and letting you correct it, and to redirect when a nervous candidate overextends. A chatbot's instinct is the opposite, to fill every pause and agree, so it has to be told.

Feed it two more things. STAR, as the UK's National Careers Service defines it: the situation, the task you were given, the action you took, and the result and what you learned. Their worked example ends on a number, a 25% increase in newsletter sign-ups over three months, and that is the bar for a result.

If the job is in the UK public sector, the ad will name behaviours from the Civil Service Success Profiles framework, updated in January 2025, such as making effective decisions, delivering at pace, and communicating and influencing. Paste the named ones in and make it score against those.

2. Rehearse the pay conversation

Run this before the first screening call, because plenty of them open by asking what you are expecting. The whole thing turns on holding your number through four seconds of silence, which is why it is worth saying out loud to something that will push.

PromptRehearse the pay conversation
You are the hiring manager, and you are trying to get me for as little as the budget allows, politely. Play it straight: be warm, be reasonable, and push. I want to practise holding a number while someone I like is being nice to me, because that is the part I get wrong.

FILL THIS IN FIRST
- The role and employer: [name]
- The posted range, if there is one: [paste it exactly, or "none posted"]
- What I currently earn: [amount, or "prefer not to say, and I want to practise not saying it"]
- My number: [the figure I want]
- My walk-away number: [the figure below which I would decline]
- What else matters to me: [notice period, hybrid days, holiday, training budget, title, start date, anything]
- Where my number came from: [my own research, a recruiter, a friend in the role, a guess]

IF I HAVE NOT GIVEN YOU A WALK-AWAY NUMBER, ASK FOR IT AND DO NOT START WITHOUT IT. Negotiating without one is how people talk themselves down in the room.

BEFORE WE START, ON WHERE MY NUMBER CAME FROM
If I said it came from a guess, say so plainly and send me to find a real figure first: the employer's own posted ranges for similar roles, a published pay scale if this sector has one, a union or professional body rate, or a recruiter working that market. Do not give me a figure from memory. If you can search the web, cite only an actual posted range or a published scale, with the link and the date. Anything else is "[FIND THIS YOURSELF]".

THEN RUN THE CONVERSATION, ONE TURN AT A TIME

Open the way a hiring manager does, and use the moves they actually use. Across the rehearsal, put at least four of these to me:
- Asking what I am currently earning, early and casually.
- Asking for my expectations before naming a range.
- Naming a figure at the bottom of the range as though it were the range.
- The pause after I say my number. Actually leave it. Say nothing but "[silence]" and wait for me to reply.
- "That is above what we budgeted for this role."
- "We could get you there at the next review."
- Enthusiasm about everything except the money.
- Asking me to confirm I would accept the lower figure right now, on the call.

Stay in character. Do not coach me mid-conversation. If I fill the silence by lowering my own number, let it happen and note it.

THEN DEBRIEF ME
- Every point where I moved, quoted back to me, with what I said just before I moved.
- Whether I ever named a figure below my stated number, and how quickly.
- Whether I gave away my current salary, and what I could have said instead.
- What I asked for besides money, and what I left on the table.
- One specific line to use next time for each moment I fumbled, in my own register rather than a script.

NEVER
- Never quote a salary benchmark, a market average, or a percentage rise from memory. Cite a posted range or a published scale with a link, or write "[FIND THIS YOURSELF]".
- Never tell me my number is too high or too low. You do not know.
- Never let me end the rehearsal without having said my number out loud at least once without qualifying it.

3. The debrief, within an hour of the real one

Do this while you can still hear the questions. GitLab's handbook makes the payoff explicit: where hiring is structured that way, later interviewers cannot see the earlier ones' full feedback, deliberately, to avoid bias. So a story that landed in round one can be used again, and a question that caught you out will very likely come round again with someone new.

PromptThe debrief
I have just come out of a real interview and I want to capture it before I forget it. Your job is to get it out of my head accurately, then turn it into something I can use for the next round. Do not comfort me and do not tell me how it went. You were not there.

ASK ME THESE ONE AT A TIME, IN THESE WORDS, AND WAIT AFTER EACH

1. "List every question you can remember, in any order, in their words as closely as you can."
2. "Which one did you answer worst, and what did you actually say?"
3. "Which one did you answer best, and what made it land?"
4. "What did they seem most interested in? What did they come back to, or write down?"
5. "What did they tell you about the role, the team or the process that was not in the ad?"
6. "What did you promise to send or follow up on?"
7. "What did you not get to say?"

Keep going until I say I am done remembering. If an answer is thin, ask one follow-up, then move on.

THEN BUILD ME FOUR THINGS

A. THE QUESTION LOG. Every question, grouped by what it was assessing. Mark the ones likely to come up again with a different interviewer.

B. THE TWO FIXES. The two answers to rewrite before the next stage, with what was missing in each: a specific situation, my own action rather than the team's, a result, or a number. Use only what I have told you. Anything I need to supply, write as [I NEED TO ADD THIS].

C. THE REUSE LIST. The stories that worked. Where a process runs several separate interviews, a later interviewer often has not seen the earlier feedback in full, so a story that landed can be used again with someone new.

D. THE FOLLOW-UP. A short thank-you note from what I actually said, referencing one specific thing from the conversation, and anything I promised to send. Keep it under 150 words and write it in my register, not in yours.

NEVER
- Never add a detail about the conversation that I did not give you.
- Never guess at whether I got it. If I ask, say you have no way of knowing and go back to the fixes.
- Never invent a fact about the company to put in the follow-up note.

FINISH WITH
The single thing to practise before the next round, and one sentence on why that one.

The honest bit

Practising with it and using it in the room are different things, and the clearest statement of that comes from the company that makes Claude. Anthropic's published candidate guidance, last updated in July 2025, tells applicants to use Claude to research the company, practise answers and prepare questions, then says of live interviews that this is all you, no AI assistance unless they indicate otherwise, because they want to see how you think in real time. Take-home exercises, the same. Not every employer agrees: Final Round AI sells a feature it calls Stealth Mode for use during live interviews, so check the policy of the place you are applying to. The likelier risk sits in your own preparation: over-rehearse and the answer sounds recited, so use these to tighten your own words and leave the script out of it.

Do the mock tonight, badly

Take the actual ad, paste it into the first prompt, and let it interview you before you are ready, because the scores from a cold run are the only honest ones you will get. Then spend an hour on GitLab's interviewer handbook, read as the candidate, because knowing they are filling in a scorecard changes what you say.

Optional: Yan Liu's free Claude skill turns your real experiences into a tagged bank of STAR stories and leaves the results blank for you to fill in, so it cannot invent them. It carries no licence file, so learn from it rather than build on it.

If the CV is the part still holding you up, tailoring it to the actual ad comes first.

A few quick questions

Can I do this by typing instead of talking?

Yes, and the marking still works. What you lose is the silence and the rambling, the two things the pay rehearsal and the redirect exist to train, so if you only type, read your answers aloud as you go.

What if the job ad names no competencies?

Write none named in that field and let the prompt derive them from the ad, quoting the line it took each from. That quoting step is the check: if it cannot point at a line, it is guessing at what the job assesses.

How many times should I run the mock?

Three runs. Once cold, again after you have rewritten the three weakest answers, then a third time in the two days before the interview on the questions you scored lowest. Beyond that you start sounding recited.