Prompt Engineering for Recruiting: Screen Resumes, Run Interviews, Draft Offers
Hiring is one of the highest-leverage workflows you can hand to a language model — and one of the easiest to get wrong with a lazy prompt. This playbook gives you field-tested prompt patterns for every stage of recruiting: screening resumes, generating interview questions, comparing candidates, and drafting offers. Copy the templates, adjust the variables, and ship a hiring pipeline that stays fast and fair.
Why recruiting breaks generic prompts
Most teams fail at AI-assisted hiring because they treat it like a chat task: "look at this resume and tell me if it's good." That prompt has three fatal flaws. First, there is no rubric — the model invents criteria each run, so the same resume can get two different verdicts. Second, there is no job description in context — without it, "good" means whatever the model guesses. Third, there is no output format — you get a paragraph of vibes instead of a score you can compare across candidates.
The fix is prompt engineering discipline: always bind the model to a job description, a scoring rubric, and a structured output. Every prompt in this guide follows that pattern. If you want the general theory first, read our prompt engineering playbook before continuing.
Resume screening prompts that actually work
A resume screen is a classification task: strong fit, possible fit, or no fit — with evidence. Paste the job description and the resume into the prompt, then force the model to cite its evidence. The template below rejects resumes that fail hard requirements and scores the rest across four dimensions.
Why this pattern beats "is this resume good?"
- Hard requirements first cut false positives before scoring noise creeps in.
- Quoted evidence makes every screen auditable — you can trace any verdict back to a resume line.
- Structured JSON lets you compare 200 candidates in a spreadsheet instead of reading 200 paragraphs. See our guide to structured output prompting for the full technique.
Generate interview questions by role level
Copy-paste interview questions are useless — a senior engineer and a junior designer need different probes. The trick is to prompt the model with the level, the competencies, and the format of the interview round, then ask for questions that cannot be answered with a generic anecdote.
This turns a 30-minute question-writing chore into a 5-minute task, and the "weak answer" notes act as an alignment sheet for interviewers who haven't done the role before.
Compare candidates with a shared rubric
The most expensive recruiting mistake is comparing candidates on different criteria — interviewer A judged "culture fit," interviewer B judged "systems thinking," and the debrief becomes an argument about vibes. Prompt the model to normalize raw interview notes into one shared rubric before the debrief.
| Stage | Prompt input | Model output |
|---|---|---|
| Normalize | Interviewer notes + rubric | Each note mapped to a rubric dimension, 1-5 score |
| Compare | All normalized scorecards + JD | Rank-ordered table with strengths, risks, open questions |
| Calibrate | Scorecard + past calibration data | Flag scores that deviate from historical hiring bars |
Run this on every candidate and your debriefs get shorter, calmer, and anchored to evidence instead of memory.
Draft offers and rejections that sound human
The last mile of recruiting is writing — offer letters, rejection emails, reference requests — and it's where most AI drafts fail by sounding robotic. The fix is to prompt with voice constraints: who is speaking, who is receiving, and what must never be said.
What never to automate fully
- Final hiring decisions — the model advises; a human signs.
- Salary negotiation language — review every number against policy.
- Reference checks — real conversations, not summarized ones.
Anti-bias and compliance guardrails
Hiring prompts are a compliance surface. Build these guardrails into every template, not as afterthoughts:
- Strip identifiers first. Redact names, photos, addresses, graduation years, and personal interests before the resume ever reaches the model.
- Ban proxy signals. Explicitly forbid the model from inferring age, gender, ethnicity, disability, or family status — and from using proxies like "cultural fit" without a definition.
- Keep the human in the loop. Every AI screen, score, and draft is a recommendation. Log prompts, outputs, and the human decision for every candidate.
- Check local law. Some jurisdictions (including New York City and the EU under the AI Act) regulate automated employment decision tools. Know your obligations before deploying.
Pair these prompts with the right tooling from our AI tool recommendations for HR and recruiting teams to cover the parts prompts alone can't — scheduling, sourcing, and background checks.
Limits and notes
These prompts assume clean inputs: a real job description, real resumes, real interview notes. Garbage in means confidently wrong scores — audit the first 20 outputs against a human reviewer before trusting the pipeline. Models also drift: re-calibrate rubrics quarterly, version your prompt templates (see prompt version control for AI teams), and never feed candidate data into a model that trains on inputs without a data processing agreement.