A hiring manager sends you three lines: "Need a good backend person, 3 to 5 years, should be a team player. Can you post it today?" By evening you'll have a job ad full of adjectives, a pile of resumes of wildly different shapes, and a manager who says "none of these feel right." AI prompts for HR and recruiters help most at two ends of that chain: turning a vague request into a clear job description and rubric, and turning scoring evidence into fair, documented notes.
They help least in the middle, where a model is asked to decide who is good. This post is built around that split. Models draft; people judge. Rejections are never automated.
These are untested templates; I haven't run them in a live hiring process. The worked example is invented. Nothing here is legal advice. Hiring and data-protection rules vary by country and sometimes by state, so check your policy with your legal or HR leadership. The wider prompts for HR post covers onboarding, policy and performance; this one stays on recruiting.
How do I write a job description with AI that doesn't read like every other ad?
Feed it the work, not the title. The first prompt turns the manager's three lines into questions you take back to them.
A hiring manager gave me this request: "[PASTE THEIR WORDS]"
Do not write the job description yet. Produce:
1. Ten questions I should ask the manager to turn this into a clear role,
covering: what the person will own in the first 90 days, what success
looks like at 6 months, who they work with, the tools and stack, what
is genuinely required vs learnable on the job, location and working
pattern, and what the last person in the role struggled with.
2. For each question, say what a vague answer would look like, so I can
push back.
Then, with the answers, write the ad.
Write a job description for [ROLE TITLE] at [COMPANY, one line on what it does].
FACTS (use only these):
- Outcomes in first 90 days: [LIST]
- Day-to-day work: [LIST]
- Must-have skills (max 5, each tied to a task): [LIST]
- Nice-to-have: [LIST]
- Team and reporting line: [DETAIL]
- Location / working pattern: [DETAIL]
- Pay range and benefits: [ONLY IF YOU HAVE APPROVED NUMBERS; else write
PAY RANGE: TO BE ADDED]
- Hiring steps: [LIST]
RULES:
- Open with what the person will do, not "we are looking for".
- Plain language. No "rockstar", "ninja", "fast-paced environment",
"culture fit" or similar filler.
- Do not add requirements that are not in the facts, including degree or
years-of-experience minimums unless I listed them.
- Do not invent benefits, perks or numbers.
- End with a short, honest "what this role is not" line.
Two checks before posting. First, read the must-haves and ask whether each one blocks a good candidate for a reason unrelated to the work. Years of experience and degree filters often do. Second, ask the model to review its own draft:
Review this job description for wording that may discourage qualified
applicants (for example gendered wording, unnecessary credentials, vague
culture phrases) and for requirements that are not clearly tied to the
duties. List each issue with the exact phrase and a plainer alternative.
Do not claim the ad is free of bias.
The last sentence matters. A model can catch some phrases. It cannot certify an ad as fair.
How do I build a screening rubric before I see any resumes?
Write the criteria first. A rubric made after you've read resumes tends to describe whoever impressed you.
Create a screening rubric for [ROLE] from this job description:
[PASTE]
Output a table with: Criterion | Why it matters for the work | Evidence to
look for in a resume or application | Weak / Adequate / Strong descriptors.
Maximum 6 criteria, each job-related and observable.
RULES:
- Do not use criteria based on name, age, gender, marital status,
nationality, religion, caste, photo, or where someone studied as a proxy
for quality.
- Treat career gaps as a question to ask, not a negative score.
- Include a column "Ask in interview if unclear".
- State assumptions you made about the role.
Get the manager to approve the rubric in writing. Then it becomes the reference for every candidate, which is the single best protection you have against drift.
Can AI summarise resumes against the rubric?
It can, with strict limits. Summarising evidence is different from ranking people. The prompt asks for evidence and gaps, not a verdict.
Rubric: [PASTE]
Candidate application text (personal identifiers removed: no name, contact
details, photo, date of birth, address):
[PASTE]
For each rubric criterion, quote the exact lines from the application that
provide evidence, or write "NO EVIDENCE FOUND". Then list questions to ask
this candidate.
RULES:
- Do not give a hire / reject recommendation.
- Do not give a numerical overall score.
- Do not infer age, gender, background or personality.
- If the text is unclear or in an unusual format, say so.
The output goes to a human reviewer who decides who moves forward. Every candidate who doesn't advance should be reviewed by a person against the rubric, not filtered out because a model summary looked thin. "NO EVIDENCE FOUND" often means the resume omitted something, not that the person lacks it.
Privacy first. Resumes are personal data. Check your policy and the tool's data terms, strip identifiers, and prefer tools your employer has approved. The post on what not to paste into ChatGPT is a useful checklist. If your team is building a screening pipeline rather than pasting, the counterfactual approach in testing LLM bias with swapped names shows how to test for different outputs on otherwise identical inputs.
What do I put in an interview kit?
A kit means every candidate for the role gets the same core questions and is scored against the same descriptors.
Create a structured interview kit for [ROLE] using this rubric:
[PASTE]
For each criterion give:
- Two questions: one behavioural ("Tell me about a time...") and one
situational or practical for this role.
- What a strong, adequate and weak answer would contain (concrete signals,
not adjectives).
- A follow-up probe.
Also provide a scoring sheet: Criterion | Score 1-4 | Evidence noted
(candidate's words) | Interviewer initials.
RULES:
- No questions about age, marital or family status, pregnancy, health,
religion, caste, nationality, or other personal characteristics.
- Questions must be answerable from the candidate's own work experience
or a task.
- Do not invent facts about our company.
Ask the interviewer to write evidence before scoring, and to score before the debrief. Those two habits do more for consistency than any model.
How do I write candidate replies, including rejections?
This is a good use, because the failure mode is a cold or vague message.
Write a reply to a candidate for [ROLE] who [STAGE: applied / interviewed
once / interviewed twice] and will not be moving forward.
Facts I can share: [e.g. "we chose candidates whose recent work was closer
to X"]. Facts I cannot share: [e.g. comparison with other candidates].
Under 120 words. Warm, respectful, specific about the stage. No false
promises ("we'll keep your CV on file") unless I confirm that is true:
[YES/NO]. Offer feedback only if [YES/NO].
Keep the reason honest and general. A human sends it, and a person made the decision.
A worked example (illustrative)
An invented scenario: the manager asks for "a good backend person, team player, 3 to 5 years." After the question prompt, the manager says that in the first 90 days the hire will take over an internal billing service, add one report and cut incident pages. Required: comfortable reading someone else's code and writing tests. Python helps but is learnable.
The rubric (shape):
| Criterion | Evidence to look for | Ask if unclear |
|---|---|---|
| Takes over existing systems | Maintained or inherited services, not only greenfield | "What was the last system you inherited?" |
| Testing habit | Tests written, how they were used | "Describe a bug your tests caught." |
| Production support | On-call or incident work | "Walk me through your last incident." |
| Python or fast learning | Python use, or a quick switch of stack | "What did you pick up fastest?" |
Notice what's gone: "team player", a degree filter, and "3 to 5 years" as a gate. The rubric can still screen on the work. A person with six years or two years can show the same evidence. Whether you keep an experience range is a decision for the hiring manager and your policy, not a default.
Failure modes to check for
| Failure | What it looks like | Fix |
|---|---|---|
| Inflated requirements | Ten must-haves, years and degree filters | Cap at five; tie each to a task |
| Auto-rejection | Model output used to drop candidates | Human review of every non-advance |
| Biased proxies | College name, gaps or location treated as quality | Forbidden-criteria rule; review rubric |
| Fabricated evidence | A skill "found" that isn't in the resume | Quote-only rule; read the original |
| Identifiers pasted | Names, phone numbers, photos in a chat | Strip before use; policy check |
| Perk invention | Benefits or pay the company never offered | Facts-only rule; leave pay as a placeholder |
| Different questions per candidate | Hard to compare, hard to defend | One kit per role |
What not to automate
- The decision to reject. A person reads the rubric evidence and decides, and records why.
- Ranking candidates by a model score. Scores look precise and aren't. Use evidence summaries.
- Anything inferring personal traits. No guessing age, gender, background or personality.
- Legal and compliance wording. Offer terms, notice periods, background checks and data retention need HR or legal sign-off.
- Candidate data in unapproved tools. Remove identifiers or don't use it.
- Sensitive conversations. Disability adjustments, complaints and references are for people.
For writing and structuring prompts that stick to supplied facts, avoiding hallucinations is the underlying pattern. If you also write candidate-facing documents, the prompt library has a job description template and a performance review template you can adapt.
