How to Prompt AI to Write a Hiring Rubric

To prompt AI to write a hiring rubric, tell it the job title, the four to six competencies the role actually requires, and one explicit rule: no vague traits like "team player" or "culture fit," only skills you can observe in an interview answer or work sample. Ask the model for a 1-5 scoring scale with a one-sentence behavioral anchor for each score, tied to that specific competency, formatted as a table every interviewer can use the same way. Below is the exact prompt template, a blind-scoring step that cuts bias before it starts, and the legal reason a human still needs to sign off on the final decision.

Why a Generic AI Prompt Produces a Weak Rubric

Ask a chatbot to "make me a hiring rubric" and it typically returns a list like communication, teamwork, culture fit, and leadership potential. Those categories feel reasonable, but none of them are tied to what the job actually requires, and vague categories are exactly where interviewer bias creeps in. Two people can watch the same answer and score "culture fit" in opposite directions for reasons that have nothing to do with the job.

Structured interviews built on defined, job-specific competencies and a consistent rating scale are associated with less biased and more predictive hiring outcomes, according to guidance from the Society for Human Resource Management (SHRM). The fix is not a smarter AI model. It is a better prompt that forces the model to name specific, observable, job-relevant criteria before it ever writes a score.

A Prompt Template That Forces Role-Specific Criteria

Use this as your starting ai hiring rubric template. Replace the bracketed section with your role, and paste the actual job responsibilities underneath before you send it to the model.

You are helping design a structured interview scorecard for the role of [JOB TITLE]. Using the responsibilities below, identify 4 to 6 competencies that are directly required to do this job well. Do not include generic traits such as culture fit, team player, or leadership potential unless you can define them as an observable, job-specific behavior. For each competency, write a 1-5 scoring scale where 1 describes a response that shows the competency is missing or is a red flag, 3 describes an adequate but unremarkable response, and 5 describes a response that clearly exceeds what the role requires. Output the result as a table with columns: Competency, Score 1, Score 3, Score 5. Then list two interview questions or work-sample tasks per competency that would let an interviewer actually observe it. [Paste job responsibilities here]

Three parts of this prompt do the real work. Banning vague trait language outright removes the category most likely to get scored on gut feeling instead of evidence. Requiring a 1, 3, and 5 anchor for every competency forces the model, and the interviewer using its output, to describe what each score actually looks like instead of guessing on the fly. Asking for questions or work samples tied to each competency means the rubric is not just a scoring sheet, it also shapes what gets asked in the room.

This works because it applies the same core prompt engineering habits that make any structured output reliable: be specific about the format, constrain the vocabulary, and require the model to justify each output with evidence rather than adjectives.

If you are hiring for the same type of role repeatedly, it is worth turning this into a standing tool rather than retyping it every time. The same logic used to write a system prompt for a custom AI assistant applies here: fix the rules, no vague traits, mandatory 1-5 anchors, evidence-based questions, once, and let only the job title and responsibilities change each time you run it. Pairing this with a prompt that writes the job posting itself, such as one that helps you prompt AI to write a job description, keeps the responsibilities and the rubric describing the same job in the same language.

Turn the Output Into a Structured Interview Scorecard

Here is what that prompt produces for one competency on a senior data analyst rubric: SQL and data modeling. Use it as a reference for how specific your own anchors should be.

Every interviewer scoring against this table is judging the same thing: could this candidate do the SQL and modeling work the job requires. There is no row for "seemed sharp" or "would fit in well." If a trait cannot be pinned to a 1, 3, and 5 description like this, it does not belong in the rubric.

Prompt AI to Screen Candidates Fairly With a Blind First Pass

A rubric only reduces bias if it gets used before an interviewer knows who the candidate is, not after. Ask candidates for written answers to the same two or three job-specific questions your prompt generated, or a short work sample, and score those responses against the rubric with names, photos, and school names redacted. Only unmask the candidate and their full resume once every criterion already has a 1-5 score attached to it.

That rubric-first, resume-second sequencing is what actually blocks bias: once a reviewer has seen a name, a school, or a photo, the score they were about to give tends to bend toward what they already believe about that candidate, even when they are trying not to let it happen.

Blind review is not a complete fix on its own. Research on blind hiring has found it helps most at the specific stage where it is applied, and can be undercut by other signals, like a zip code or a listed hobby, that hint at identity even after a name is removed. Treat it as one layer in the process, not the whole solution.

An AI Rubric Supports the Hiring Decision, It Does Not Make It

Everything above produces a consistent scoring framework and a fairer first pass. It does not make the final call, and it should not. Automated tools used to screen or score candidates can trigger separate legal obligations depending on where you hire. New York City's Local Law 144, for example, requires many automated employment decision tools to pass an independent bias audit and requires candidate notice before use, and Title VII of the Civil Rights Act still applies to any selection process, automated or not, that produces a disparate impact on a protected group, regardless of intent.

Keep a human reviewer who can see the full picture, resume, work history, references, and the interview itself, and let the rubric structure that person's judgment instead of replacing it.

That human review step matters even more once you consider that not every resume you screen is what it appears to be. Before the rubric stage, it is worth knowing how to spot an AI-generated resume when hiring, since a scorecard built on a fabricated work history is only as reliable as the resume behind it.

Frequently Asked Questions

What should I include in an AI hiring rubric prompt?

Include the job title, four to six job-specific competencies pulled from the actual responsibilities, an explicit ban on vague trait language, a 1-5 scoring scale with a behavioral anchor for each score, and a request for interview questions or work samples tied to each competency.

Can AI legally screen job candidates?

AI can help structure and score candidate evaluations, but automated screening that meaningfully affects a hiring decision can trigger bias-audit and notice requirements in jurisdictions such as New York City, and remains subject to anti-discrimination law everywhere else. Use it to support, not replace, human judgment.

How do you stop AI from writing a vague hiring rubric with things like culture fit?

Give it an explicit instruction banning generic traits unless they can be defined as an observable behavior, require a 1-5 anchor for every competency, and tie each criterion directly to the job responsibilities you paste into the prompt.

What is a blind first pass in hiring?

A blind first pass means scoring a candidate's written answers or work sample against the rubric before seeing their name, resume, or other identifying details, then revealing their identity only after every score has already been recorded.

How many competencies should a hiring rubric have?

Four to six is a workable range for most roles. More than that becomes hard for interviewers to score consistently in a single conversation, and fewer risks missing something the job actually requires.