Give the Model the Standard Before Asking for Feedback

Why Does the Model Say Every Answer Is Good?

Because a question like "is this a good answer" hands the model no standard, so it works from the only material it has, which is your answer. Ask for an opinion and an opinion comes back. Ask instead which parts of a named requirement the answer fails to cover, and the reply has something outside your own text to check against.

The praise is not a bug in the model so much as an answer to the question you actually asked.

The usual fix circulating in prep guides is a wording change: tell the model to be brutally honest, or to act as a skeptical interviewer. That does shift the tone. It rarely adds information, because the model still has nothing to compare your answer to, and harsher language wrapped around the same empty judgment is not a critique. You end up with "this could be stronger" in a sterner voice, which tells you no more than the compliment did.

The fix sits upstream of the wording, in what the prompt contains.

Two things change the output. One is the standard you supply. The other is what you ask the model to produce: a list of gaps can be checked line by line, and a verdict cannot.

Give It the Posting and the Competency

Paste three things before you ask for anything: the job posting, the question exactly as it was asked, and your answer as you would say it out loud. Then name the competency you think the question is testing, and ask the model to list what a hiring team would find missing against that posting. The instruction is "what is not here", not "how did I do".

There is a real standard behind this, and you do not have to invent it. The US Office of Personnel Management's guide to structured interviews describes a format where every candidate gets the same questions in the same order and is rated on a common scale. The interviewers also agree in advance on what an acceptable answer looks like ¹. That guide is federal HR practice from 2008 and binds no private employer. It still describes the format a structured panel works from.

That agreed answer is the thing your practice partner has been missing.

So write the standard down yourself before the model sees it. Four inputs are enough, and the fourth is the easiest to leave out.

• The posting, pasted whole, including the responsibilities nobody reads twice
• The question as asked, in the interviewer's words rather than your summary
• Your answer, transcribed from saying it aloud, not written out neatly
• The competency you believe the question targets, named in one phrase

Then ask for one thing: the specific claims in your answer a hiring team could not verify from it, with the missing detail named for each. That output is a to-do list, and the length of it is the useful part.

Naming the competency does more work than it looks like. Buehl and colleagues had 111 students answer the same structured interview twice, once honestly and once told to answer as an applicant, and how far each person's rating rose in the applicant run tracked how well they could identify the dimensions being assessed ⁵. Those were psychology students in a research interview with no job at stake, and the rise the study measured is the effect of that instruction, not a practice gain. The transferable part is narrow and useful: knowing what a question is for changes what you put in the answer.

Name it wrong and the critique goes looking in the wrong place.

If you cannot tell which competency a question targets, the posting usually says so somewhere below the summary. Where it stays vague, using AI for company research and then checking three sources gets you closer to what the role actually rewards.

Ask for the Follow-Ups, Then Answer Them

Once the gaps are listed, ask the model for the three follow-up questions an interviewer would ask next, and then answer them out loud. The ones you stall on are the holes in the example, and they are the same holes a second question would have found in the room. This is the part of practice a chat window can actually simulate.

Employers are given a matching ladder from the other side. The hiring-side guidance on what follow-up questions expose whether someone understands the answer they gave runs three rungs. Specify which alternative you ruled out, invert the constraint, then say what would have to be true for the answer to be wrong.

A rehearsed monologue survives the first question and rarely survives the third.

Some live AI interviewers now work this way. Alex markets an AI recruiter that runs live phone and video interviews, and its own candidate-testimonial page describes the AI summarizing a candidate's answers back and asking follow-up questions based on them ². That is a vendor describing its own product on a testimonial page, so read it as a claim rather than a verified account, and the mechanics may vary by employer. It still tells you what to prepare for.

Prepare for a conversation rather than a set of speeches.

That ladder is published guidance for interviewers, not a secret. Reading what the employer side is told to ask is a faster way to see what your answer is being asked to prove.

Where the Model's Judgment Runs Out

A model can find a missing number, an unnamed constraint or a story with no decision in it. What it cannot tell you is whether the example you picked matters for this role. In the Harvard Business School field experiment, consultants worked through a set of tasks with GPT-4, and one of those tasks was chosen to sit outside what the model handled well.

On that task, the consultants using the model were 19 percentage points less likely to reach the right answer ⁴. That is one task in one 2023 study, on a model that is no longer current, and it is not evidence that a model makes anyone worse in general. The finding worth carrying is narrower. On that evidence, the consultants could not tell which side of the line the task was on.

So take the content to a person who has actually hired for the role.

A former manager, a friend two years ahead of you, anyone who has sat on the panel. They can tell you whether the example is the right size for the job, which is the judgment the model was guessing at.

Practicing this way is inside the rules at least one employer publishes. Anthropic's candidate guidance, dated July 2025, tells applicants to draft the written application themselves and then use Claude to refine it, to use it freely for interview preparation, and to do live interviews with no AI assistance unless the employer says otherwise ³. That is one company's policy about its own product, open to revision, not a norm.

Where nothing is published, ask the recruiter what the guidance is for this process. It is a fair question and a short one. The applicant-side piece on whether practicing interview answers with ChatGPT counts as cheating covers where that line usually sits.

Nothing here is a way around a rule an employer has already written down.

Judging your candidacy was never on the list. The model is a fast way to find out what your answer leaves out, which is the part you can still fix before the call.

See how it works >