Role-Specific Interview Question Generator

You are producing a packet the interviewer reads before the conversation: the question, the competency it tests, and what a strong versus weak answer sounds like. You are not producing a quiz the candidate never sees scored.

Speed is the use case. A project role opens, you need a specialist on a workstream in days, and nobody has an afternoon to invent a fair interview from a blank page. A general LLM such as Copilot or ChatGPT will fill a kit in one sitting if you paste the role. The cost of that speed is a tidy list that may include illegal questions, a recycled strategy case, and trivia the job will never use.

The candidate hears the questions in the room. The strong and weak signals stay with the interviewer so two people can be compared on the same competencies. If you later record a rating in an interview kit in Greenhouse or Ashby, that rating is a human judgment after the conversation. Do not ask the model to grade candidates.

This is also not a skills-to-project matching engine. Matching ranks who might fit. This page is what you ask once they are on a shortlist.

Three inputs change the questions. A generic job description produces generic questions.

1. The project role as it will be staffed: workstream, duration, client setting, who they sit with, what they must leave behind when the engagement ends.
2. A short competency list. Six named competencies is plenty. Twenty competencies produce twenty shallow questions.
3. What good looked like on a similar past project. A few sentences on the behaviors that distinguished the person who actually delivered, not the person with the shiniest deck.

"Good" is the prompt that makes this role-specific. Without it, the model fills from consulting-interview folklore: leadership stories, market-sizing cases, and textbook definitions.

Do not dump the last person's CV. Dump the work. They sat with the client's analyst until the joins were trusted. They refused to calculate a number on untrusted data. They could explain a price change to a sales director without a formula dump. They taught the client to run the file after week twelve.

Interest and motivation are a different signal. Use a consultant preference and interest matcher for whether someone wants the work. Do not turn the interview into a personality quiz.

This is an illustrative scenario, not a case study and not reported results.

A staffing lead needs a workstream lead for a twelve-week pricing transformation at a mid-market consumer-goods client. One consultant plus one analyst, sitting with commercial finance. Named competencies: reconciling messy SKU and promo data; facilitating skeptical finance stakeholders; designing a simple price-pack structure; leaving the client able to run the analysis without the team.

What good looked like last year on a similar job: the strong person documented every join, delayed elasticity work until the file was trusted, and could walk a sales director through a pack-size change in plain language. The weak person built a model nobody used.

A usable generated item

Question: "On a past pricing or margin project, walk me through the first week when the client's SKU file did not match finance's margin report. Who did you sit with, and what did you refuse to calculate until the join was trusted?"

Testing: data reconciliation under ambiguity, not tool trivia.

Strong signal: names the mismatch, the client counterpart, the documentation habit, and a decision they delayed on purpose. Weak signal: "I cleaned the data in Python" with no stakeholders, or a lecture on joins with no client mess.

A second item, more technical, still tied to the workstream:

Question: "How would you structure a price-pack analysis when the hierarchy is incomplete and promotions overlay list price? What do you ask the client before you build anything?"

Testing: method for this workstream. Strong signal: sequencing, the questions they would ask, and what they will not compute yet. Weak signal: recites a textbook elasticity formula and starts drawing a model.

Generate a follow-up probe with each question so the interviewer does not go shallow after the first sentence. "What did you show the client, and what did they still not believe?" is more useful than a second unrelated prompt.

What unconstrained generation dumps into the same kit

Left to "write interview questions for a consulting data lead," the same prompt often adds a market-sizing case about umbrellas, "what is the formula for price elasticity," "tell me about a time you showed leadership," and, worse, "culture" probes about background, family, or where someone is "really from."

None of those test the workstream. The case is campus lore. The formula is trivia. The leadership prompt is interchangeable across every role in the firm. The culture probes can be illegal.

A hiring or staffing lead reviews and cuts the set. The model does not ship to interviewers unedited.

Illegal and discriminatory questions. Models will generate questions about age, family plans, pregnancy, health, religion, nationality, visa status, disability, or "culture fit" that is a proxy for background. Those are not interview content. Hiring rules differ by country. Asking the same job-relevant questions of every candidate on a slate is often a legal duty, not a preference. When a question could be a protected-characteristic probe, delete it. If you have employment counsel, they see the reviewed kit, not the raw chat.

Generic McKinsey-case clones. Market sizing, paper profitability cases, and "how would you enter this market" unless that is the actual job. You are staffing a project role or hiring someone onto a live engagement. Campus case interviews are a different process. A pricing workstream lead who can crack a paper case and freeze in a messy client extract is the wrong person for this slot.

Trivia that does not appear in the role. Definitions, formulas, and tool quizzes. Knowing a textbook formula by heart does not mean they can run a client workshop. If the competency is "reconcile messy commercial data," ask about that week, not about a definition.

Keep the set short. Six questions asked with probes beat twenty asked quickly. Every surviving question should map to a named competency and to a sentence in the "what good looked like" note. If it cannot, cut it.

Copilot or ChatGPT drafts. Greenhouse or Ashby interview kits, or your equivalent scorecard, are where a structured set usually lives so two interviewers are not inventing different hours. That is a filing choice, not a product ranking, and it is not a reason to skip the human cut. If your kit is a shared document, that is enough at this effort level. This is a prompt-and-review workflow, not a platform build.

Same questions across candidates on a slate is what makes comparison possible. Changing the kit mid-slate because the model offered a "better" question is how you lose the comparison.

Do not paste candidate CVs into the generator to "personalize" questions unless you have a clear policy for that. Personalization is how you drift into protected characteristics and into trivia scraped from a public profile.

Do not treat the scorecard as an auto-grader. The interviewer still listens, still probes, and still records a judgment. The generated strong and weak signals are a calibration aid, not a hidden answer key.

The kit ends when you have a staffing decision. Onboarding is a different packet.

Once the person is on the project, a project setup automation agent, a kickoff deck auto-assembler, and a stakeholder onboarding brief synthesizer are the next artifacts. Do not keep interviewing in week one. Do not recycle interview questions as kickoff content.

The system is working when, the day a role opens, interviewers share one short guide tied to this project's competencies, and a human has already removed the illegal question, the case clone, and the trivia. It is failing when each partner improvises, or when a generated "culture" question makes it into the room.