How to Write Interview Questions Using ChatGPT for Any Role

Why AI Interview Questions Sound Generic

ChatGPT produces clichés when you give it a job title and nothing else. With no duties, no constraints and no definition of a good answer, it does the only thing available to it — averages every interview question on the internet and hands you the middle of that pile.

Watch what changes when the input carries some weight.

The second question can only be answered properly by someone who has done the work. And notice where the quality came from: not clever phrasing, but constraint. Telling the model what to rule out reliably does more work here than telling it what to include, which is the same principle behind negative constraints in prompting. Three inputs do nearly all of the work:

• The actual duties, pasted verbatim out of the job description
• A named ban on the openers you never want to hear again
• A definition of what a strong answer to this question contains

Most hiring managers stop one step short of this. They add the duties but never tell the model what a strong answer contains — so it writes questions that sound specific and still can’t be scored.

Ask any talent acquisition director what actually goes wrong in hiring and you will rarely hear “our questions were bad”. You will hear that three interviewers came out of the same room with three different verdicts and no way to reconcile them. That is a scoring failure wearing a question-writing costume, and it is the reason the rest of this guide spends as much time on the answer key as on the questions.

The Answer-Key Method Explained

The Answer-Key Method is one rule: pin the competency, write the answer you would score a 5, then reverse-engineer the question and the probes that would surface it. The question gets written last.

This is not a trick we invented. The US Office of Personnel Management’s summary of the interview research describes the most structured interviews as ones where every applicant is asked the same pre-defined lead and probe questions and scored against benchmarks of proficiency — and reports that more structured interviews show higher validity, better agreement between raters, and less adverse impact. It also notes that questions built on competencies identified through job analysis hold up best of all. In practice that means three things:

• The same lead and probe questions go to every candidate
• Answers are scored against written benchmarks of proficiency
• Competencies come from the job itself, not the job title

Which is the entire argument for doing it this way. The Answer-Key Method we teach at PromptPeakAI is just that research turned into six prompts you can run on a Tuesday afternoon before a Wednesday interview.

It also explains why so much advice on how to write interview questions using ChatGPT produces disappointing results. Those guides optimise the wrong artefact. A beautifully phrased question attached to no benchmark is still an unscoreable question, and a plain question attached to a clear benchmark will out-hire it every time.

How to Write Interview Questions Using ChatGPT

Here is the working version of how to write interview questions using ChatGPT: four steps, run as one conversation. Keep them in a single chat so each prompt can see the output of the last — the model needs that continuity to keep the questions tied to the competencies you agreed on in step one.

Paste the job description and ask for the short list of things the interview must actually test. Four to six is the working range. Fewer and you can’t distinguish candidates; more and a 45-minute interview can’t cover them.

You are an experienced hiring manager helping me design a structured interview.

Read the job description below and identify the 4 to 6 competencies this interview must test. Split them into technical skills and behavioural skills. For each one, write a single sentence explaining why it matters for performance in THIS role specifically.

Do not add competencies that are not supported by the job description. Do not include generic corporate values.

Job description:
[paste the full job description]

Read the output as a hiring manager, not a proofreader. If a competency doesn’t map to something the person will do in their first quarter, strike it out and say so in the chat.

Here is where this method departs from every prompt listicle on the internet. Before any question exists, you define what a great answer contains.

For each competency in the list above, write a benchmark answer profile with three parts:

1. STRONG (score 5): the specific evidence, ownership language and numbers a genuinely strong candidate would mention.
2. ADEQUATE (score 3): what a competent but unremarkable answer contains.
3. WEAK (score 1): the vague or borrowed-credit answers that signal the person has not done this work.

Describe observable behaviours and evidence, not personality traits. Keep each part to two sentences.

This output is the most valuable thing in the session. It becomes your questions, your probes and your scorecard — and it takes about ninety seconds to produce.

Now, and only now, the questions. You want two types: behavioural questions about what the candidate has already done, and situational questions about what they would do. Both formats work; the research suggests behavioural questions hold up especially well for complex professional and managerial roles.

Using the benchmark answers above, write one behavioural interview question per competency.

Rules:
- Each question must be answerable ONLY by someone who has personally done the work.
- Ban generic openers: no "tell me about yourself", no "greatest weakness", no "difficult customer" phrasing.
- After each question, list 3 follow-up probes I can use if the answer is vague. Probe 1 targets their individual contribution, probe 2 targets the numbers, probe 3 targets what went wrong and what they changed.

Format as a numbered list. Keep every question under 30 words.

Now write 2 situational scenario questions for the same role.

Each scenario should describe a realistic problem this person will actually face here, based on the job description. End each with: "walk me through your first three moves."

Constraints:
- Only reference tools and systems named in the job description. Do not assume or invent any software.
- For each scenario, tell me what an experienced candidate covers that an inexperienced one misses.

That “do not assume or invent any software” line matters more than it looks. Left alone, the model fills gaps with whatever tools are statistically common for the job title, and you end up asking a candidate about a platform your team has never bought.

If the questions read stiffly, that is fixable in one line: ask ChatGPT to rewrite them in spoken English, the way you would actually say them out loud. The structure underneath stays intact. This is the same Role, Goal, Context, Format discipline behind the Logic System — you are correcting Format, not starting over.

Turn the Answer Key Into a Scorecard

A scorecard is what stops three interviewers walking out of the same conversation with three different verdicts. And because you already wrote the answer key in step two, the rubric is mostly assembly rather than invention.

Turn the benchmark answers into a candidate scorecard.

For each question, build a 5-point scale where 1 = Unsatisfactory, 3 = Proficient, 5 = Exceptional. Write a one-sentence descriptor for levels 1, 3 and 5 describing what the candidate SAID or DID at that level — never what they seemed like.

Add a final column called "Evidence noted" for the interviewer to fill in during the conversation.

Output as a table I can paste into a document.

The difference between a scale that works and one that doesn’t is whether the levels are anchored in observable behaviour.

Here is the shape of what comes back for a single competency.

Print it. Give every panel member the same copy. The scorecard is the artefact that turns four separate opinions into one comparable set of scores.

Adapt the Kit for Any Role

The four steps never change. What changes is how much you feed into step one and how deep the probes need to go — and getting that calibration right is most of the difference between a kit that works for a warehouse supervisor and one that works for a finance director.

• Technical and specialist roles. Name your actual stack in the prompt and ask for a five-minute live scenario rather than a take-home. Engineering managers will tell you that watching someone reason out loud beats any exercise they can complete unsupervised.
• Senior and executive roles. Cut to three or four competencies and go deeper on each. Add one prompt asking for a question about a decision the candidate later reversed — seniority shows in how people describe being wrong.
• High-volume or entry-level hiring. Keep the kit short and identical for every candidate. Consistency is doing more work for fairness here than sophistication ever will.
• First-time interviewers on the panel. The probe ladder matters more than the questions. Give a nervous interviewer three pre-written drill-downs and they stop accepting the first pleasant answer they hear.

One habit worth building regardless of role: run the whole thing as a multi-turn conversation rather than trying to extract everything from one giant prompt. Layered prompts that each build on the last consistently beat a single request for questions, probes, scenarios and a rubric all at once, and they let you correct the competencies before the errors compound downstream.

What You Can Safely Paste In

Everything above involves pasting company material into a chat window, so this section is not optional reading. The short version: role material is usually fine, people material is not.

The rule most articles get wrong is which line actually matters. It is not free versus paid. On OpenAI’s published plan comparison, every personal plan — Free, Go, Plus and Pro alike — carries the same entry: content is used to train the models, with an opt-out available. Paying twenty dollars a month changes your message limits, not that default.

One practical note on long documents: personal accounts have a modest context window, so a fifteen-page requisition may quietly get truncated. Paste the responsibilities and requirements sections rather than the whole PDF. Our guide to using ChatGPT safely at work covers the settings in more detail.

Audit the Questions Before You Ask

You are about to use these questions to make a decision about someone’s livelihood, so run one more pass. This one takes thirty seconds and catches the two things that get organisations into genuine trouble: questions that stray into protected territory, and questions that quietly favour a particular background.

Review every question and probe in this interview kit as an employment compliance reviewer.

Flag any question that:
- touches age, family status, health, disability, religion, nationality, or anything else outside the job requirements
- assumes a specific career path, education route or life circumstance the role does not require
- tests confidence or polish rather than the competency it claims to test

For each flag, explain the risk in one line and give me a corrected version that tests the same competency.

What ChatGPT Cannot Judge

The honest limits of this workflow sit in one place: it can build the instrument, but it cannot conduct the interview or make the call.

There is also a symmetry worth sitting with. The candidate on the other side of the table is running the same tools — building a company brief, predicting your questions and rehearsing structured answers, much as we describe in our guide to preparing for a job interview using AI.

That is not a problem for this method; it is the argument for it. Rehearsal survives a generic question and dies on the third probe, because no amount of preparation invents a metric you never owned.

When the Output Goes Wrong

Four failure modes account for nearly everything that goes sideways here, and each has a one-line fix you can paste straight into the same chat.

That second row is the one to watch. Invented tools, invented certifications and invented industry standards are the classic symptom of AI hallucination, and in an interview they are unusually costly — you look unprepared to the candidate, and you may screen out someone for lacking a requirement that was never real.

Keep the kit for next time

Once a role’s kit works, it is worth keeping. The competencies change per role but the four steps never do, which makes this an obvious candidate for a Custom GPT preloaded with your competency framework — available from the Go tier upward. The next hiring manager on your team then gets a consistent interview kit without reading any of this.

The scorecard also earns its keep a second time. The evidence you recorded against each competency is exactly what makes a specific, humane rejection email possible instead of a form letter. So run this once on the role sitting on your desk right now — knowing how to write interview questions using ChatGPT costs about fifteen minutes, and the version of you walking into that room on Wednesday will be far better prepared than the one who improvised in the corridor.