Operations

Drafting Job Postings and Interview Questions with AI: Screen for Decisions, Not Duties

Can an AI assistant write your job posting? Yes — and the first draft will probably read fine. The problem is what it screens for. A posting assembled from a duties list attracts people who can do tasks. A posting built around the decisions the role actually owns attracts people who can be trusted with them. The fix is not a cleverer prompt. It is a different input: three real decisions, stated before the AI writes a word.

Published 2026-10-27 · Last updated 2026-10-27

The answer first: draft the decisions, not the duties

Start here, before any tool. Name three real decisions the person in this role will own in their first 90 days, and write down what a wrong call on each one costs the business. Not responsibilities — decisions. “Responsible for inventory” is a duty. “Decides when to reorder stock and how much, with a stockout costing a weekend of sales” is a decision with a price attached.

Those three decisions become the spine of everything the AI drafts for you. The posting screens for them. The interview questions probe them. The scoring rubric grades answers against them. Change the input this way and the same general-purpose assistant — the ChatGPT, Claude, or Gemini class of tool — produces a hiring document instead of a form letter. Keep the duties list and it will produce a very polished version of the posting that has already failed you.

The posting: inputs the AI actually needs

A posting prompt is only as good as its inputs. The set that produces a usable draft is specific: job title, business type, location, hours, salary range, three to five responsibilities, and two separate skills lists — must-have and nice-to-have, kept separate. Add the three decisions from the section above, and set two constraints on the output: keep it under 400 words, and no corporate jargon.

Two of those inputs do more work than the rest. The salary range, because a posting without one asks applicants to guess and asks you to negotiate blind later. And the split skills lists, because a single merged list teaches the AI — and the reader — that everything matters equally, which means nothing is actually required. If a skill is a must-have, it belongs in the screening. If it is a nice-to-have, it belongs in the interview, not the filter.

The procedure: posting to interview pack in one sitting

1. Write the three decisions first. Ninety-day horizon, one line each, with the cost of a wrong call beside each. This is your work, done before the AI sees anything.
2. Feed the full input set. Title, business type, location, hours, salary range, three to five responsibilities, must-have skills, nice-to-have skills, the three decisions. Ask for a draft under 400 words with no corporate jargon.
3. Read the draft against the decisions. Does the posting actually screen for them, or did the decisions get smoothed into duties on the way through? Send it back with the decision wording restored until the posting names what the role owns.
4. Generate the interview pack as a separate deliverable. Specify it exactly: 8 questions, each with a written description of what a strong answer sounds like and what a weak answer sounds like. At least two of the eight must ask about a decision the candidate got wrong.
5. Ask for one scoring rubric. One rubric, applied to every candidate for the role — the same questions, the same strong/weak standards, the same scoring. That consistency is the entire point of writing the standards down in advance.
6. Run the bias pass (next section) on the posting and the questions before anything is used.
7. Approve it yourself before it goes anywhere. Human approval before the output is published is a named guardrail, not a courtesy. The draft does not post itself.

What success looks like: you can pick any interview question in the pack and point to the decision it tests, the strong answer it should produce, and the weak answer it should expose — and every candidate for the role faces the same eight questions scored on the same rubric.

Why the strong/weak pairs matter more than the questions

A question on its own is half a tool. “Tell me about a time you handled a difficult customer” sounds like an interview question, but without a standard attached, what it really measures is how the answer strikes whoever is listening that day. The strong/weak description is what converts the question into an instrument: a strong answer names the decision the candidate made, the options they weighed, and the result; a weak answer describes what the team did, or what should have happened, in the passive voice.

The two wrong-decision questions carry the most signal. Asking about a decision that went badly — and listening for whether the candidate can name their own part in it — tests the exact thing the posting screened for: judgment, not task completion. A candidate who has never made a wrong call has either never owned a decision or will not tell you about it. Both answers tell you something.

Store the finished pack — posting, questions, rubric — where the hiring actually happens. If your business already uses an applicant tracking system such as Manatal or Workable, the rubric belongs inside it, attached to the role, so the next hire starts from the same standard instead of a blank page.

The bias pass, as a checklist

“Check for bias” is usually offered as a slogan. Run it as a procedure instead. A bias-checker prompt follows a specific pattern: it flags demographic signals in the text, rewrites biased language, and attaches a skills-focused rationale for each change — so you can see not just what was changed, but why, in terms of the skills the role requires.

• Use explicit instructions. The bias check only happens if the prompt asks for it directly. Vague requests get vague passes.
• Read every flagged phrase yourself. The tool flags and proposes; you decide. A rewrite that removes a genuine requirement is as much a failure as a flag that was missed.
• Check the questions, not just the posting. A clean posting followed by interview questions that probe availability, background, or personal circumstances in ways unrelated to the three decisions has failed the same test at a later stage.
• Set rollback criteria in advance. Decide before you publish what would make you pull the posting and redraft — a flag you cannot resolve, a requirement you cannot tie to the work. Guardrails named in advance get used; guardrails improvised mid-problem get negotiated away.
• Approve before publishing. Nothing the AI drafted goes live until a person has read the final version. This is the human-in-the-loop step, and it is the last item on the checklist because it is the one that catches everything the others missed.

What the owner still decides alone

The framework, plainly: the AI drafts, screens language, and formats. It does not decide what the role is. The three decisions, the must-have line, the salary range, the standard for a strong answer — those are judgments about your business that no prompt can supply, because they are the content the prompt runs on. An owner who skips them gets fluent text wrapped around a vacancy. An owner who writes them gets a posting that filters, an interview that measures, and a rubric that makes two candidates comparable on paper instead of in memory.

That is the whole division of labour. Decisions in, documents out.

General information only — not legal, employment, or human-resources advice. Hiring and human-rights obligations vary by jurisdiction; if you are unsure whether a posting or question is permissible where you operate, have it reviewed by a qualified professional before you publish.

— Dex Carter, Tech Writer