AI Interview Prep Prompts: Turn ChatGPT or Claude Into a Mock Interviewer

You have a panel interview on Thursday. You've read the job description four times and you've rehearsed answers in the shower, but you've never been asked a follow-up by anyone who wasn't on your side. That's the gap a good AI mock interviewer fills.

The trick is that the default chatbot is a bad interviewer. Ask it to "interview me" and you get five questions in one message, then a gushing review. This post gives you a setup prompt that fixes that, a rubric to score against, role variants, and a way to catch the model being too nice.

The setup prompt for a mock interviewer

Paste this at the start of a fresh chat. Fill the brackets, then paste your CV and the job description below it.

You are a hiring manager interviewing me for the role of [ROLE] at [COMPANY or
"a company like X"]. You are fair but not friendly. You've interviewed many
candidates and don't hand out credit for vague answers.

Rules:
1. Ask ONE question at a time. Then stop and wait for my answer.
2. After my answer, ask one follow-up that probes the weakest part of what I
   said (a number I didn't give, a decision I didn't justify, "what was your
   specific role"). Only then move on.
3. Do not give feedback during the interview. Stay in character.
4. Run [8] questions: 2 about my background, 4 behavioral, 2 role-specific.
5. When I type "SCORECARD", stop and score each answer using the rubric below.

RUBRIC (score each answer 1-5):
- Relevance: did I answer the question that was asked?
- Specificity: concrete situation, numbers, names of tools or decisions
- Ownership: clear on what I did, not "we"
- Result: stated outcome and what I learned
- Concision: under about 2 minutes spoken
For every score, quote the exact sentence from my answer that justifies it.
A 5 requires all of the above. Do not round up.

Start with your first question.

MY CV:
[paste]

JOB DESCRIPTION:
[paste]

Why these rules? Rule 1 stops the five-questions-at-once habit. Rule 2 is where the value is: real interviewers dig into your weakest sentence, and models won't do it unless told. Rule 3 keeps the practice honest, because feedback mid-interview makes you perform for the feedback. The "quote the exact sentence" line forces the model to ground its score in what you actually said, which cuts down on generic praise.

Role-playing prompts like this work because you're giving the model a persona with incentives. The role prompting playbook goes deeper on why, and the assigning roles lesson covers the basics.

What should the mock interviewer ask? Role variants

Swap the first paragraph of the setup prompt. Everything else can stay.

For software roles there's a separate guide to prompting as an Indian software engineer if you want prompts for the on-the-job side of the career as well.

How do I get STAR feedback that's actually useful?

STAR stands for Situation, Task, Action, Result. It's a way to structure behavioral answers so they have a beginning, a middle and a measurable end. Most weak answers fail in the same three places: they spend 80 percent of the time on Situation, say "we" instead of "I" during Action, and skip Result entirely.

After you finish a session, run this on any single answer:

Here is my answer to "[QUESTION]":
"[PASTE YOUR ANSWER OR TRANSCRIPT]"

1. Mark which words belong to Situation, Task, Action and Result.
2. Tell me which of the four is missing or thin.
3. Flag every "we" that should be "I", and every claim that has no number.
4. Rewrite the answer in 150 words or fewer, using only facts I stated. Do not
   invent details. Where a number or fact is missing, put [NEED: ...] so I can
   fill it in.

The last line is the one that matters. Without it, models "improve" your answer by inventing a 30 percent cost saving that never happened. If you then say that figure in the room and someone asks how you measured it, you're finished. Only polish what's true.

Here's the shape of the output you want. It is illustrative, not a captured transcript.

Your answer: "We had a lot of bugs after launch and I helped the team fix them
and things got better."

Feedback: Situation is vague (how many bugs? what was affected?). Task is
missing. "I helped the team" hides your action. Result "got better" has no
measure.
Rewrite: "After the [NEED: product/feature] launch, [NEED: number] bugs were
reported in the first [NEED: period]. I was responsible for [NEED: your part].
I [NEED: what you did], and [NEED: outcome with a number]."

That rewrite looks like homework, and it is. That's the point. It shows exactly what you need to remember before Thursday.

Make the scorecard trustworthy

Models tend to be agreeable, so you should assume the scorecard is inflated until you've tested it. Two checks that take five minutes:

The bad-answer test. In a throwaway chat with the same setup, give the model a deliberately terrible answer ("Um, I'm a hard worker and a team player, I guess my biggest weakness is perfectionism"). If it scores that above 2, tighten the rubric: add "A generic answer with no specific example scores 1 for Specificity" and try again.

The re-run test. Score the same answer in two fresh chats. If the scores differ by more than a point, don't trust the numbers. Trust the quoted sentences and the missing-element notes. Use scores for trends across sessions, not as truth.

One more habit: ask it to name the single most damaging sentence in your answer. It's easier for a model to find one concrete problem than to be calibrated across five scores.

Questions the model won't ask unless you tell it to

A model interviewing you from a CV tends to produce safe, generic questions. Add one of these to the setup when you want a harder session.

• "Ask about the gap, the short tenure or the career change that a skeptical interviewer would raise."
• "Ask one question where the honest answer is 'I don't know'. See how I handle it."
• "Ask me to estimate something (users, cost, time) and challenge my assumptions."
• "Ask what I'd do differently on the project I'm most proud of."
• "Ask why I'm leaving, and push back once if my answer blames my current employer."

And the one every candidate forgets: ask the model to roleplay the end of the interview, where you ask the questions. "Now it's my turn. Respond as the hiring manager to the questions I ask, and tell me afterwards which of my questions showed research and which could have been answered by the careers page."

What AI can't do for you here

• It can't hear you. In text, you can edit and polish. In a real interview you ramble. If your chatbot offers a voice conversation mode, use it for at least a few rounds, and check your plan's limits because they vary. Alternatively, record yourself answering, transcribe it and paste the transcript into the STAR prompt.
• It doesn't know this company's actual questions. It will guess from the job description. For real signal, search for the company's interview process on sites where candidates post experiences, and treat those as anecdotes.
• It can't judge fit or presence. Warmth, energy, how you handle silence: get a friend to run one session with you.
• It will invent if you let it. Never let it add facts to your answers. Keep the [NEED: ...] rule.
• Privacy. Strip phone numbers, addresses, referee names and anything under NDA from the CV or work examples before you paste. If you're discussing a previous employer's internal numbers, you shouldn't be pasting them anywhere.

A 3-day plan

1. Day 1. Run the full 8-question session. Type SCORECARD. Save the two lowest-scoring answers.
2. Day 2. Rewrite those answers with the STAR prompt, fill in the [NEED] gaps from memory, then run a second session with a different persona (stress or panel).
3. Day 3. Run a short 4-question session on your weakest competency only. Then do the "my turn" round to prepare your own questions.

For ready-made prompts on other work tasks, see the prompt library. The HR prompts guide shows what recruiters ask AI to help them screen, which is worth knowing before you phrase your answers.