How to Use AI to Prepare for a Job Interview

What I found: I asked ChatGPT to predict twelve interview questions from a job posting and to label each one DIRECT if it came from the posting’s own wording. Nine came back DIRECT. I checked all nine against the posting line by line: six held up, one was a generic opener that would fit any job on earth, and two were built from the background I had pasted rather than the posting. The question list is worth an hour of your time. The labels on it are not worth trusting.

If you are in a hurry: paste the posting and your own history into an AI chat, ask for the questions an interviewer would actually ask, and use it as a sparring partner rather than a scriptwriter. The trap is letting it write answers for you — it will invent achievements you never had, and you will not notice until you are saying one out loud.

What the list actually looked like

A ChatGPT table of likely interview questions, each tagged DIRECT or INFERRED with a note on what the interviewer is checking — The first five of twelve questions, with the source tag and the “what they’re really checking” column. ChatGPT, September 2026.

Nine of the twelve came back tagged DIRECT. I checked each one against the posting word for word, and six of them held up — they traced to real phrases like “the patience to chase suppliers”, “weekly production planning with the roast team” and “the person who notices when a shipment is going to be late”. The other three did not. One was a generic opener that would fit any job on earth, and two were built out of the background I pasted rather than the posting. The sorting is useful; the labelling is not something to lean on.

What this test could not tell me

The posting and the candidate were invented — a fictional operations coordinator role at a fictional coffee roastery — so that I could check the output against something I fully controlled. That is also the limit of it. Nobody sat the interview, so I cannot tell you how many of the twelve questions a real interviewer would have asked. What I can tell you is which of the model’s own DIRECT labels survived a line-by-line check against the posting, and that is the part nobody checks.

The rest of this is how to run it yourself.

Give it the three things it cannot guess

Most bad interview prep comes from asking a cold question like "what are common interview questions for an operations job". You get the same generic list that is on every careers blog. The model only becomes useful once it has material that is specific to you and this role.

Before you ask anything, paste in three things:

The full job posting, copied from the listing rather than summarised. The phrasing matters — interviewers often build questions straight out of the bullet points they wrote.
Your own history in rough form. Bullet points are fine. Real projects, real numbers if you have them, the parts you would rather not be asked about.
What kind of interview it is — a screening call with a recruiter, a panel, a technical round, a final with the hiring manager. The questions differ enormously and the model will default to a generic first-round if you do not say.

If the posting is confidential or the company is small enough that the details identify you, strip the company name before pasting. That is a judgement call worth making deliberately; I wrote about where the line sits in the guide on pasting work documents into AI tools.

Turn the posting into a question list

The first useful ask is narrow. Not "give me interview questions" but something closer to this:

Here is the posting and my background. Give me twelve questions this interviewer is likely to ask. For each one, tell me what they are really checking. Mark each question as either DIRECT if it comes from specific wording in the posting, or INFERRED if you are guessing from the role type.

That last sentence is the important one. It gives you something checkable. You can read the posting and confirm whether a question tagged DIRECT really traces back to a line in it, which tells you immediately how much of the list is grounded and how much is filler.

Let it structure your material, never supply it

This is the step where the whole exercise can quietly go wrong, and the failure is hard to spot. If you ask it to "write a strong answer" to a question, it will produce a polished, confident paragraph containing a specific accomplishment. That accomplishment will sound exactly like something you might have done. It will not be something you did.

In my session I asked for a sample answer to a question about handling a supply shortage, and it came back with a story about renegotiating terms with a secondary supplier and cutting stockouts by a specific percentage. I had given it none of that. It was a plausible shape filled in with invented detail, and if I had been tired and reading fast I might have half-absorbed it as my own.

The version that works is the reverse. Give it your raw material and let it do the structuring:

Here is roughly what happened in that project, in my own messy words. Turn it into a two-minute spoken answer with a clear situation, what I specifically decided, and the outcome. Do not add any facts I have not given you. If something is missing, ask me instead of filling it in.

That last instruction does real work. It will come back with questions like "what was the actual timeline" rather than inventing one. Answer those, and the result is your story, tightened.

The same principle applies to how the answer reads. If the draft sounds nothing like you, the fix is to feed it your own writing and ask it to match the register — the method in the guide on getting ChatGPT to write in your own voice works just as well for spoken answers.

Rehearse against it, out loud

Reading prepared answers on a screen builds almost no useful muscle. Saying them does. The most valuable twenty minutes of the whole exercise is this prompt:

Interview me. Ask one question, wait for my answer, then give me one line of feedback on the answer before moving to the next. Be direct about padding, rambling, and answers that never actually answer the question.

Then answer out loud first and type a compressed version of what you said. The gap between the fluent thing you typed and the thing that came out of your mouth is the whole point of rehearsing.

The feedback quality here was uneven. On rambling and on answers that dodged the question, it was sharp and specific. On anything requiring judgement about the industry, it was confidently generic — "consider quantifying the impact" is advice that fits every answer ever given and helps with none of them.

Prepare the questions you will ask them

Almost every interview ends with "do you have any questions for us", and almost everyone wastes it. This is the one part of interview prep where the model is straightforwardly good, since it is reading the posting more carefully than you are.

Ask it: Based on this posting, what is unclear, unusual, or possibly a warning sign? Give me five questions that would surface it without sounding suspicious.

In my run it noticed that the posting listed both bookkeeping and wholesale account management under one coordinator role, and suggested asking how those two halves had been split before — which is a polite way of asking whether the job is secretly two jobs. That is a useful observation, and not one I had made myself.

The parts worth being careful about

Three things to keep in view.

It cannot tell you about this company. Anything it says about the culture, the team, the interviewer or recent news is either from stale training data or invented outright. Use a search engine and the company's own site for that, and treat any specific claim from the chat as unverified. The habit of checking is worth building generally; I covered how in the guide on fact-checking an AI answer.

Memorised answers sound memorised. The goal of all this is to know your three or four stories cold enough to tell them differently every time, not to recite a paragraph. If you find yourself learning sentences, you have gone too far.

Do not paste anything covered by an NDA — and remember that a detailed enough project description identifies a client even without the name on it.

If you only have forty minutes

Paste the posting and your background. Ask for the twelve questions with what each one is really checking. Pick the four that scare you most. For each, tell it what actually happened in your own words and have it tighten the telling without adding anything. Then run the mock interview on just those four, out loud. That is most of the value of the full exercise, and it fits in the evening before.

Search This Blog