Interview question generator: AI interview questions from a job description, scored on your rubric

Pull the competencies out of the description first, then write questions against those competencies rather than against the job title. A good generator does this in two passes: it reads the description to identify the four to six capabilities the role actually depends on, ranks them by how much they drive performance, then writes two to four behavioral or scenario questions per capability.

The ranking step is what separates a useful output from a generic one. Most job descriptions list fifteen requirements, and maybe five of them predict anything. Cangrade's free tool is explicit about this: it returns five soft skills ranked by impact and two questions per skill, which is a sensible shape. A generator that hands you twenty questions has skipped the prioritizing and pushed that work back onto you.

The failure case to watch for is keyword echo. If your description says "must be detail oriented" and the generator returns "tell me about a time you were detail oriented", it has paraphrased the requirement rather than designed a way to test it. A question that works asks about a situation where the trait would have to show up, so the candidate has to describe events rather than agree with an adjective.

The same competency map is what your scoring should hang off later, which is why generating questions and building the rubric are really one job. We go deeper on that on our structured interview software page.

Five to seven questions is the realistic ceiling for a 30 minute screen, and six is a comfortable landing spot once you allow for follow-ups. A behavioral question with a real answer and one probe takes about three minutes. Budget five minutes at the start for context and role framing and another five at the end for the candidate's own questions, and you have roughly 20 minutes of actual question time.

For a 45 minute conversation the number moves to about eight or nine, and for a full hour, ten to twelve. Those are ceilings rather than targets. Interviewers who try to force twelve questions into 30 minutes end up cutting every answer short, which produces the worst of both worlds: you asked a lot and learned very little, because nobody got past the rehearsed opening of any answer.

This is the constraint that most generated lists ignore. A tool that returns ten questions for a 30 minute screen has handed you something you cannot run as written, and the interviewer will silently drop four of them. Which four varies by interviewer, and that variance is exactly the thing structure is supposed to remove.

Two to four per competency is the standard guidance, and two is usually enough at the screening stage. One question per competency is fragile, because a candidate can miss a single question for reasons that have nothing to do with capability, including simply not having a good example ready. A second question gives them another route to demonstrate the same thing.

Past four you hit diminishing returns and start eating the time you needed for other competencies. If a capability genuinely warrants six questions, it is usually two capabilities wearing one label, and splitting it will make your scoring clearer.

At first-round screening specifically, three or four competencies with two questions each is a sound structure. Deeper capability testing belongs in the later stages where a hiring manager is in the room and has an hour rather than half of one.

The practical answer to what interview questions to ask a candidate is narrower than most generated lists suggest: ask about things the candidate has actually done, in situations similar to the ones the role will put them in. Behavioral questions that start with "walk me through" or "describe a time when" outperform hypotheticals, because a hypothetical measures how well someone imagines and a behavioral question measures what they did. Both are easy to write badly, so anchor each one to a competency you named.

Keep every question job-related. That is not only better signal, it is the safer position legally: questions that stray into age, family status, national origin, disability, religion or arrest history create real exposure, and a generated list can wander there without meaning to. If you want the specifics, our guide to illegal interview questions covers what to strike.

Ask every candidate on the requisition the same questions in the same order. This is the least glamorous rule and the one that does most of the work, because it is the only thing that makes two candidates comparable. It is also the rule that breaks first when a human is running the fifth interview of the day.

A question is scorable when someone has written down what each rating level sounds like before the interview happens. That means anchors: a description of a level 2 answer, a level 3, a level 4, phrased in terms of what the candidate says rather than how the interviewer feels. Without anchors, a 1 to 5 scale collects impressions and averages them, which looks like measurement and is not.

Most generators stop one step short of this. They give you the question and sometimes a short note on what to look for, which is genuinely helpful, but the note is usually a single sentence describing a good answer. That tells an interviewer how to recognize a 5 and leaves the entire middle of the scale undefined, which is where nearly every real candidate lands.

Use five levels with a written description of each, and expect most scores to be 3s. Reviewers cluster in the middle when the scale is vague, so a well-anchored 3 is a genuine finding rather than a shrug. The mechanics of building the anchors and keeping them consistent across interviewers are covered on candidate interview scoring.

For producing a first draft of a question list, yes, and there is no reason to pay for that step alone. We ran four of them on August 31, 2026 and the questions were serviceable. If your problem is a blank page at 9pm before a 10am screen, a free generator solves it completely and you should use one.

They are not sufficient when your problem is consistency across candidates, defensibility, or volume. A free generator produces a document. It does not ask the questions, does not record who was asked what, does not score anything, and leaves no evidence trail if a rejected candidate or a regulator later asks how the decision was reached. Those are the parts that cost money to solve, which is why they are the parts nobody gives away.

It is also worth noticing what the free generators are for. GoHire and Cangrade both sell hiring platforms, and the generator is the top of that funnel. That is a fair trade and the tools are useful, but it means the generator is deliberately scoped to end where the paid product begins. We compared what each one returns, side by side, in our roundup of the best interview question generators.

Yes, and this is the part current models are genuinely good at. Given a real job description with duties and requirements in it, a model will produce questions that are specific to the role rather than generic, and it will do it faster and more consistently than most people writing from scratch on a Tuesday afternoon.

The quality tracks the input closely. A thin description of four bullet points produces generic questions, because there is nothing specific to work from. Paste the full description including the day-to-day responsibilities, the tools involved and what the first six months look like, and the output changes noticeably.

Where it still needs a human is judgment about what matters. A model cannot know that your last two hires failed on stakeholder pushback rather than technical depth, or that this particular team needs someone comfortable without a manual. Give it that context explicitly and it will build questions around it. Leave it out and you get a competent interview for a generic version of your role.

The hiring manager should define the competencies and the recruiter should own the question set and the process. The manager knows what the role actually fails on. The recruiter knows what can be asked legally, what fits in the time, and what has worked across other requisitions, and is the one who has to keep the process consistent across every candidate.

A generator is a good forcing function for that conversation, because it produces a concrete draft in two minutes and arguing with a draft is much faster than starting from nothing. The useful meeting is fifteen minutes spent cutting a generated list of twelve down to six and writing anchors for those six, not an hour spent staring at a blank template.

Whoever writes them, lock the set before the first interview. Questions that get edited between candidate two and candidate seven quietly destroy comparability, and it is the sort of thing nobody notices until a hiring decision needs defending.

The questions themselves carry the ordinary risk of any interview question, which is that a non-job-related question about a protected characteristic is a problem regardless of who wrote it. Review a generated list before you use it. A model working from a job description that mentions "recent graduates" can produce questions that edge toward age, without any intent to.

The bigger regulatory question is about scoring rather than generation. If software issues a score or a recommendation that substantially assists an employment decision, rules like New York City Local Law 144 apply, which means an independent bias audit within the previous twelve months, a published summary, and ten business days of notice to candidates. Generating a question list does not trigger that. Scoring the answers can.

The duty follows the job rather than your headquarters, so a New York requisition brings New York obligations even if your office is in Denver. Illinois and Maryland have their own notice and consent rules. We keep the current state of this on our AI hiring compliance page.