Practise a job interview with AI

The interview is on Thursday at ten, and you want this one. Nerves can make even a familiar question difficult to answer. It is "so, why are you leaving?" or "tell me about a project that went wrong". You have answers to both. You have never once said either of them out loud.

Rehearsing in your head does not fix that, because your head accepts a half-finished sentence and moves on. A room does not. What an AI model is genuinely good at here is narrow and useful: it will ask you a question and then wait, thirty times in a row, without getting bored, and you can be terrible at it in private as many times as you like.

Here is the run, from the job ad to the questions you ask them at the end. Twenty-five minutes of this beats reading a list of a hundred common interview questions.

1. Give it the ad, and give it you

Open a new chat and paste the job ad in full. Under it, paste eight to ten lines about yourself: your current role, how long you have done it, and three things you actually did, with the numbers you would be willing to defend if someone pushed.

Both halves matter. The ad on its own produces questions written for a job title, which come out as "tell me about a time you demonstrated leadership". Your background on its own produces questions for a different job. Together you get the awkward ones: the four-month gap, the sideways move, the tool the ad lists twice that you have used exactly once. (Rewriting your CV against the same ad is a separate job, worth doing before this one.)

Below is a job ad and a summary of my experience. You are the hiring manager for this role. Interview me. Ask one question at a time, wait for my answer, and do not answer for me. Start with your opening question.

2. One question at a time, and mean it

The instruction people skip is "one at a time", and it is the whole exercise. Ask for a list of likely questions and you will get thirty, read all thirty, feel prepared, and walk into Thursday having practised nothing. Reading a question and answering it are different acts, and only one of them happens on Thursday.

Watch for drift. Three or four turns in, models slide back into list mode: two questions in one message, or a question with a suggested answer helpfully attached underneath. Say it again. "One question. Wait for me. Do not write my answer."

3. Answer the way you would answer in the room

Type it in one pass, in the shape you would say it, false start included. Do not tidy it up. A polished paragraph turns this into a writing exercise, and writing is not what you are being tested on.

Then say the same answer out loud to the empty room before you send it. Answers that read fine collapse when spoken, usually because the sentence is too long to finish on one breath. If your composer shows a voice call button, you can run the last pass by speaking instead of typing, against a monthly allowance of call minutes.

4. Feedback that is worth having

"How was that answer?" gets you a paragraph of nothing. Two questions get you something:

"Against the requirements in that ad, what did my answer fail to show?"
"What would you ask next to check I actually did what I just claimed?"

The second one is the good one. Real interviews are won and lost on the follow-up, and seeing the follow-up in advance tells you which part of your story you have been skating over.

Before you paste: an interview answer carries client names, unlaunched projects, sometimes numbers you signed something about. Take those out first. For the rest: we do not train on your conversations, and the provider behind a given model may. Six of the 60 flag it on their info card, 54 do not.

5. Make it a harder interviewer than it wants to be

Everything above works. The reason most people finish a mock interview feeling better than they should is that the model is a soft touch, and it is soft in a specific, predictable way.

It rates almost everything as solid. It will praise a four-minute answer that contained no example. It will not tell you that your story was boring, that you took ninety seconds to reach the point, or that the interviewer stopped listening at the second sentence, and boring is the most common reason a real interview quietly goes nowhere.

Three things change that behaviour:

Ban the praise outright. "Do not tell me what was good. List only what a hiring manager would hold against me, in order of how much it would cost me."
Put a bar under the score. "Rate that out of ten for a senior role, then tell me what the nine would have contained that mine did not." The rating will still drift high. The second half of the answer is the part you keep.
Cast it as a sceptic. "You have interviewed four strong candidates already and you are looking for a reason to say no. Push on anything vague."

Then run more than one interviewer. On any answer, branch it and choose a different model. That action is labelled Try Again on the web and Branch here in the mobile app, and it is worth reading the label before you tap: on a phone, Try again means regenerate, and regenerating deletes the answer rather than keeping it. A branch keeps everything and starts a fresh thread from that point. Friendly in one, sceptical in another, technical in a third. The threads sit separately in your sidebar and you move between them. Nothing appears next to anything else.

Here is the thing none of that fixes. The model cannot tell you whether your example is impressive for the level you are applying at. It does not know whether shipping that project in six weeks is ordinary, good or remarkable for a team your size in your industry, and it will call it strong either way, because calling things strong is what it does. That judgement is the single most valuable piece of interview feedback there is, and it comes from a person. Send your two best examples to someone who has hired for this role and ask one question: is this a good example for this level? Ten minutes of that is worth an hour of practice.

6. Practise the questions you ask them

Almost nobody rehearses this part, and the last five minutes are scored too.

Ask the model what the ad does not say that you would want to know before accepting, then make it work against you: "which of these would make me sound like I had read the ad, and which sound like I copied a list of clever questions to ask?" It handles that triage well, since it is comparing your questions against the document rather than judging you.

Then cut every question whose answer you do not actually want. Two real ones beat five performed ones.

Before Thursday

Put the context somewhere it survives the chat. Personal preferences is free, holds 5,000 characters, and applies to every new conversation: on the web it is in Settings under Account, and in the mobile app it is in your profile under Personalization, in the field labelled "Anything else the AI should know". Your role, the level you are targeting and your three strongest stories go in there once. (Memories, which you ask it to store and can organise by folder, are on the paid plans.)

What you carry into Thursday is the fourth version of the answer you kept going blank on, said out loud in your own voice, having survived someone asking "and what did you do when that failed?". The model's real contribution was patience. It asked until the question stopped sounding like a threat.

The free plan is enough to run this. Pro at $20.99/mo opens the rest of the catalogue, and you can switch models mid-conversation when you want a second interviewer.