An AI interview assistant: prompts during the interview and a broken hiring market

Knowing your craft and being able to show it are different skills

An engineer with eight years of experience cannot remember the difference between UNION and UNION ALL. He wrote it the day before yesterday. He explained it to an intern last month. Right now, twenty minutes into an interview, his head is a smooth white field, and there are five faces in small rectangles on screen, all waiting.

This is not a story about a weak specialist. It is a story about the fact that an interview tests competence and, on top of it, the ability to retrieve competence from your own head while being watched. The second is a separate skill, and it is barely correlated with the first. People who design systems for hundreds of thousands of users regularly fail a conversation about what they do every day.

That is where the demand for an AI assistant during interviews came from. Not «pass it for me», but «do not let me forget the obvious». The difference between those two things is what this article is about.

What actually breaks in an interview

The mechanism is dull and physiological. Stress narrows working memory: what is within reach when you are calm has to be dragged up with effort. On top of that, an interview asks questions in an order you never work in. At work you go from problem to solution; in an interview you are asked about the solution with no problem attached — and your brain honestly answers «it depends», because it does.

Notice that not one line on the right is about a lack of qualification. They are all about the measurement conditions. Which is why an external support looks natural: at work you have documentation, a dashboard, commit history and a colleague to ask. In an interview you have nothing, including the right to open your own notes.

Three different jobs: before, during and after

«AI for interviews» sounds like one thing and is in fact three different jobs with different value and a different cost of error. People confuse them constantly — and end up disappointed in the wrong one.

A non-obvious consequence: the most underrated part is the third. People are ready to pay for a prompt in the moment and almost never review the recording of their own interview — although that is where all the growth is. After five interviews in a row you remember impressions rather than answers, exactly as the interviewer remembers the last candidate rather than the candidates.

What «AI understands the conversation» means

The phrase is marketing, but there is concrete machinery behind it. Step by step — without it, any discussion of ethics turns into an argument about magic.

1. Speech becomes text as it goes. Not after the meeting but in real time, as a stream, split by speaker: which line is yours and which is theirs. That gives the machine what it does not have by nature — the subject of the conversation.
2. Context is added to the text. Your material: résumé, project descriptions, numbers, stack, stories. Without them a model answers in internet generalities — and that is exactly how most «AI answers» sound, the ones an interviewer recognises instantly.
3. The model sees what is on screen. A task, a diagram, a chunk of code — when the conversation is about a picture, the text of the conversation is not enough.
4. The answer arrives as support, not as a line to read. The difference is practical: a list of four points you speak from works, and a finished paragraph you read aloud does not. A person reading is audible.

AI in a technical interview: where it helps and where it hurts

The technical interview is a special case, because what is tested is not knowledge but the movement of thought out loud. And this is where a prompt hurts more often than it helps.

Live coding. Reading a finished solution off the screen is the worst possible strategy. The interviewer is not watching the code, they are watching how you get to it: which questions you ask, where you hesitate, how you check the edge cases. Somebody who silently received the answer gives themselves away instantly — they cannot explain why they chose it, and they break on «and what if the data grows a thousandfold».

System design. Here the value is real: holding a dozen components in your head, not forgetting failover and queues, checking yourself against a structure — that is work with memory, not with intelligence. Exactly what an engineer does at work by opening their own old diagram.

Factual questions. Versions, limits, exact names — the things you would look up in ten seconds at work. Keeping them in your head is a skill the profession stopped paying for.

  THEY TEST YOUR THINKING     →  a prompt gets in the way: you are audible
  THEY TEST MEMORY FOR FACTS  →  a prompt is fair: at work you would look it up
  THEY TEST EXPERIENCE        →  a prompt is useless: you either have it or not

Where help ends and deception begins

The answer depends on whose side you take — and that is not relativism, it is how the situation is built. Every participant has a different goal, and each of them calls something different «cheating».

Put the three right-hand columns together and a common denominator appears: nobody is annoyed by the tool — everybody is annoyed by the gap between what was shown and what arrived. Hence a practical rule that works in all four coordinate systems at once: use what you will use at work, and say so out loud if asked, or if you can see the question is due.

The rules becoming the norm are visible too: some companies explicitly allow AI and watch how you work with it; some forbid it at particular stages and announce that in advance; some change the format — instead of tasks with one right answer they give you somebody else's solution to critique or pair work where the dialogue is the point. All three approaches share one thing: the rule is stated. Silence is the worst option for both sides.

The mechanics of the line — what exactly substitutes a tested skill and what actually gives a candidate away — are covered in a separate piece: AI in interviews: where help ends and cheating begins. Its conclusion is enough here: if the tool stays with you at work it is help; if it is needed for exactly one hour of interview it is substitution.

An ethical view and a broken hiring market

Now the uncomfortable part, without which the ethics hang in the air. The «fair or unfair» argument is conducted as though an interview were a precise measuring instrument and the candidate the only party able to game it. Both assumptions are false.

Strong specialists are scarce, and badly found

The shortage of strong engineers is neither news nor exaggeration. The oddity is elsewhere: facing a shortage, companies make selection harder rather than easier. Five stages, a take-home task, an algorithms section, a system design discussion, a team meeting, a final with a director. At every stage there is a risk of losing a good person for a reason unrelated to the work: a bad day, a dislike of talking about themselves, no week to spend preparing for the format.

AI tools went mainstream — on both sides

Cover letters are written by a model. Résumés are tailored to the posting by a model. Screening answers too. But so is the other side: résumé filtering, candidate ranking, automated emails, sometimes a first interview stage with a bot. A situation where both sides speak through intermediaries and both pretend otherwise cannot remain ethically stable.

Recruiting automation made the first filter blind

An algorithmic filter selects on formal signals: keywords, tenure, company names, continuity of employment. By construction it cannot see the person who called the same work by different words, changed fields, or spent two years raising a child. Everyone who has looked for a job has the same experience: a hundred applications, three replies, two of them template rejections.

Take-home tasks became a cost transfer

An eight-hour unpaid task transfers the cost of selection onto the candidate, who is going through four other selections in parallel. The employer is understandable: testing by work is more reliable than testing by conversation. But when everyone does it, the total cost of job hunting grows into a second shift, and the first to drop out is the person with a current job and a child — that is, exactly the strong specialist everyone is hunting.

An interview is a poor measuring instrument, and this is known

An unstructured interview measures charm, similarity to the interviewer and the ability to talk about yourself — and only then the profession. A structured one measures better but requires preparation, shared criteria and discipline, which is why it is rarer. As a result a decision about a person is made on an hour of conversation, while the real signal is the work, which cannot be seen before hiring. This is not malice but a known limitation of the method — and it does not turn a candidate into a fraud for trying to compensate for it.

Interviewing skill has come loose from working skill

There is an entire preparation industry: task trainers, courses on answering behavioural questions, lists of standard design sections. That is direct evidence of the gap: if you can train for the measurement separately from the subject, then something other than the subject is being measured. Companies know this and carry on, because the alternatives cost more.

So the candidate plays a game whose rules already changed

They see it: an algorithm reads their résumé, a generator writes the email to them, a script checks the take-home, and in the interview they are asked about a sorting algorithm they last wrote at university. In that configuration, «please arrive with no tools at all» sounds less like ethics and more like a rule that runs one way.

It does not follow that anything goes. What follows is different: the question «use AI or not» has already been closed by the practice of both sides. The open question is which rules of use become the norm. And agreeing on them has to happen out loud, because a silent agreement is read by each side in its own favour.

AI does not cancel competence. It changes how competence is demonstrated

This is the mature conclusion the whole argument points to. The calculator did not cancel mathematics — it made testing long-division speed pointless, and exams moved to other problems. Search did not cancel erudition — it devalued memorising reference facts and raised the price of asking a precise question and telling a good answer from a bad one.

The same will happen to interviews. Testing memory for facts will become as pointless as testing an accountant's mental arithmetic. Value will shift to where the machine does not yet substitute: framing the problem, choosing among bad options, owning the decision, explaining it to people. That is good for a candidate exactly to the extent that they have something to explain.

Where Whisperer fits

Whisperer is an AI assistant for conversations: it listens, keeps a transcript, holds your context within reach and prompts support while you speak. It does not «pass the interview for you» — that phrasing is technical rather than coy: the answer arrives in a panel that you see, and the speaking is still done by a human.

Across the three jobs from the start of the article it looks like this.

Plainly about plans and limits: 60 minutes of recognition per month are free, which covers one interview with room to spare. The knowledge base, multi-screenshot analysis and the task breakdown mode are on paid plans. There is a mode in which the conversation is not stored at all: no transcript, no answers, everything deleted after the session.

And what we do not do: we do not write guides on hiding the use of an assistant, and we do not advise breaking announced interview rules. If a company asks you to come without tools, that is their right — and your right is to decide whether you want to work somewhere that states its rules honestly. Usually you do: where rules are stated, they are usually also kept.

An interview is one case of a conversation worth understanding as a whole: what an AI meeting assistant is and what it does before, during and after.

Frequently asked questions

What to do next

Testing this in a real interview is a bad idea: you will be thinking about the tool rather than the conversation. It is smarter to take your own scenario and look at it soberly.

Create an account — 60 minutes a month are free — put your résumé and a couple of project descriptions into the knowledge base and run a practice conversation: let the model ask, and answer out loud as you would in an interview. Watch three things: whether the prompt arrives when it is needed; whether it helps you speak in your own words; and whether the text left afterwards shows your weak spots.

If the answer is yes, you have a second brain that will stay with you at work too. If it is no, you will at least know precisely what gets in your way in interviews. That too is more useful than another evening with a list of standard questions.