FirstHR

ChatGPT Prompts for HR: 40 Prompts and What Not to Share

Ready-to-use ChatGPT prompts for HR tasks: job descriptions, offer letters, interviews, policies, reviews, plus what you must never paste in.

Nick Anisimov

Nick Anisimov

FirstHR Founder

Core HR
16 min

ChatGPT Prompts for HR

Forty prompts you can copy for the HR tasks a small business actually has: job descriptions, offer and rejection letters, interview questions, onboarding plans, policy drafts, review phrasing, difficult conversations, and surveys, plus the five categories of information that must never go into a general purpose AI tool

The first job description I wrote with an AI tool took four minutes and was useless. It described a person who did not exist, doing a job I had not described, in the voice of a company that was not mine. The second one took twenty minutes, because I spent sixteen of those writing the prompt, and it came back about ninety percent usable.

That ratio has held for every HR task I have handed to a general purpose model since. The output is only as specific as the input. Most pages on this subject give advice about prompting rather than prompts, which is a bit like handing somebody a book about knife skills instead of dinner.

So this is the prompts themselves, forty of them, grouped by the task you actually have in front of you. It also covers the part that gets skipped: what must never go into a general purpose AI tool, and why everything it produces is a first draft. I build HR tooling for teams without an HR department at FirstHR, and the failure I see most is not a weak prompt. It is somebody pasting a whole personnel file into a chat window to save ten minutes. This is general information, not legal advice.

TL;DR
A working HR prompt names the role the model should write as, the context only you have, the constraints it must respect, and the output format. The forty prompts below cover job descriptions, offer and rejection letters, interview questions, onboarding plans, policy drafts, review phrasing, difficult conversations, and surveys. Employee identifiers, medical information, investigation material, and named salary data never go in.

What a Working Prompt Contains

Every prompt that produces something usable contains four things: a role, context, constraints, and a format. Drop any one of them and the output degrades in a predictable way, which is why the prompts below all look structurally similar.

Definition
HR prompt
A written instruction to a general purpose language model that specifies the role it should write as, the business facts it should work from, the things it must not do, and the shape of the output. In an HR context the constraints matter more than in most other uses, because the documents involved carry legal weight and the default behavior of these tools is to fill missing detail with plausible invention rather than to ask.
Role: who the model is writing as
You are an experienced HR writer for a US small business with no HR department. One sentence. It sets vocabulary and register, and it stops the output drifting into corporate language that nobody at a twenty person company would use.
Context: the facts only you have
Headcount, industry, state, who the reader is, and what is already true today. This is the part people skip, and it is the part that decides whether the draft is ninety percent usable or a generic page you have to throw away.
Constraints: what the model may not do
Do not invent requirements I did not give you. Do not cite statutes. Do not add promises about pay or job security. Negative instructions do more work than positive ones because the default behavior of these tools is to fill gaps confidently.
Format: the shape of the answer
Six bullets, under 120 words, a two column table, placeholders in square brackets. Naming the format is the difference between a draft you can paste into a document and a wall of prose you have to reorganize by hand.
Every prompt on this page follows the same four part shape. Change the bracketed parts and the rest holds up.

The single highest leverage habit is the negative instruction. Telling a model what not to do closes the gap where it would otherwise supply confident specifics you never gave it: a benefit you do not offer, a notice period you never set, a statute that does not say what the sentence claims.

4
parts in every prompt: role, context, constraints, format
40
task specific prompts on this page
5
categories of information that never go into a chat window
1
human who reads every draft before it reaches an employee

One more framing point before the prompts. A general purpose model is a writing and structuring tool, not a decision tool. It is good at turning your rough notes into clean language and at spotting what is missing. It is bad at knowing anything about your business that you have not told it.

Job Description Prompts

Use three prompts rather than one: draft, then trim, then check the language. The draft is the least valuable step, because the model is only reorganizing what you gave it. The trimming pass is where a posting gets meaningfully better.

TaskPrompt to paste
First draft from rough notesYou are an experienced HR writer for a US small business with no HR department. Write a job description for a [ROLE] at a [INDUSTRY] company with [N] employees. These are the duties I actually need covered: [PASTE YOUR ROUGH LIST]. Give me six to eight responsibility bullets, a short must-have list, a separate nice-to-have list, and a two sentence company paragraph I will edit. Plain American English. Do not add any requirement I did not give you.
Trim inflated requirementsReview this job description and flag every requirement that is not genuinely necessary to do the job well in the first ninety days. For each one, tell me in one sentence whether it is likely to shrink my applicant pool without improving hire quality, and suggest a lower barrier alternative. Do not rewrite the posting yet. [PASTE DESCRIPTION]
Check for exclusionary wordingRead this job description and list any wording that could discourage qualified applicants or read as a preference based on age, sex, national origin, disability, religion, or family situation. Quote the exact phrase, explain the problem in one sentence, and give a neutral replacement. Change nothing else. [PASTE DESCRIPTION]
Produce a junior or senior versionRewrite this job description for a more junior version of the same role. Keep the duties that stay the same, change the scope and autonomy language, and then tell me in three bullets what actually changed so I can sanity check it. [PASTE DESCRIPTION]
Cut it down for the postingTurn this job description into a 120 word posting summary plus a one line role headline. Keep [PAY RANGE] and [LOCATION] as placeholders rather than filling them in. Do not add any selling language about culture that is not supported by the text I gave you.

The exclusionary wording pass is worth running even when you are confident. It routinely catches phrases like recent graduate, young and energetic, or a physical requirement that has nothing to do with the actual work.

Offer and Rejection Letter Prompts

Letters are the safest category for AI drafting because the content is short, the structure is conventional, and you already know every fact that belongs in it. The risk is invention: a model will happily add a term you never agreed to.

1Offer letter, first draft
PromptDraft an offer letter for a [ROLE] at a US small business. Use only these terms and change nothing else: [TITLE], [START DATE], [PAY], [PAY FREQUENCY], [MANAGER], [WORK LOCATION], [BENEFITS ELIGIBILITY DATE]. One page, warm but plain, with a signature line and a respond-by date. Do not include any term I have not given you, and do not add statements about job security, future raises, or bonus amounts.At-will language and any state specific notice requirement need a human check before this is sent.
2Rejection after a final interview
PromptWrite a rejection email to a candidate who reached the final interview for [ROLE]. Three short paragraphs, 120 words maximum. Thank them for the time they invested, say we moved forward with another candidate, and offer to keep their details for future openings. Do not give a reason, do not compare them to the person hired, and use the word unfortunately at most once.If you give a reason, give the same category of reason to everyone at that stage or do not give one at all.
3Rejection after a screen only
PromptWrite a 60 word rejection email for a candidate who completed a phone screen for [ROLE] and did not advance. Warm, specific to the role, no reason given, no promise of future contact unless I ask for one. Give me a version I can send from a shared inbox and a version I can send from my own name.Speed matters more than wording here. A same-week reply is what candidates remember.
4Declining an internal applicant
PromptWrite a message to an internal employee who applied for [ROLE] and was not selected. Account for the fact that they will work alongside the person who was hired. Be direct about the decision, name one specific development step I can commit to, and do not promise them the next opening. Under 180 words, conversational, no corporate phrasing.Deliver this in person first and send the written version afterward as a record of what you committed to.
5Reply to a counteroffer
PromptI offered [ROLE] at [AMOUNT] and the candidate asked for [AMOUNT]. Draft two replies. One holds my number and explains what else is on the table. One meets at [AMOUNT] and states plainly that this is final. Both under 150 words. Do not invent benefits, do not reference market data, and do not apologize for the number.Decide your ceiling before you read either draft, otherwise the well written one will move you.

One habit worth building: the offer letter is the document most likely to create an obligation you did not intend. Anything a model adds about tenure, promotion, or future compensation should be deleted rather than edited.

Still Using Spreadsheets for Onboarding?
Automate documents, training assignments, task management, and track onboarding progress in real time.
See How It Works

Interview Question Prompts

The best use here is not generating questions but building a structure around them: questions tied to real duties, scoring guidance written before you meet anyone, and follow-up probes for rehearsed answers.

TaskPrompt to paste
Build a structured question setWrite eight interview questions for a [ROLE] at a [INDUSTRY] small business. Four behavioral, in the form tell me about a time when. Four situational, in the form what would you do if. Every question must map to a specific duty in this job description, and say which one. For each question give me a three level scoring guide: strong answer, acceptable answer, weak answer. [PASTE DESCRIPTION]
Write follow-up probesFor each of these interview questions, write two short follow-up probes I can use when a candidate gives a rehearsed or vague answer. Keep them open rather than leading, and no more than one sentence each. [PASTE QUESTIONS]
Flag questions you should not askReview this list of interview questions and flag any that touch age, disability, medical history, national origin, religion, marital or family status, pregnancy, arrest record, or citizenship beyond basic work authorization. For each flagged question, give a lawful alternative that gets at the same job related information. [PASTE QUESTIONS]
Design a practical exerciseDesign a 45 minute practical exercise for a [ROLE] candidate that mirrors real work, can be scored by a non-expert using a rubric, and does not produce work my business would otherwise pay for. Give me the candidate instructions and a five point rubric with observable criteria.
Organize interview notesHere are my notes on three candidates for [ROLE], with names replaced by A, B and C. Organize them into a comparison against these five criteria: [LIST]. Keep my wording wherever it is evaluative, and list what I still have not learned about each person. Do not rank them and do not fill gaps with assumptions. [PASTE NOTES]

Two things about the last prompt. Replace names with letters before you paste, and delete anything a candidate told you about health, family, or immigration status rather than asking a model to ignore it. Interview notes are the most personal material most small employers generate during hiring.

The legality pass is genuinely useful, but treat it as a prompt to check yourself rather than as an answer.

Onboarding Plan Prompts

Onboarding is the task where a model earns its keep, because the work is mostly structuring: turning a list of things that must happen into a sequence with owners and dates. Give it your real task list and it will do that well.

1
Build a thirty day plan from the job description
Build a thirty day onboarding plan for a new [ROLE] using this job description. Week one is access, paperwork, and shadowing. Weeks two to four move from watching to doing. Give me two checklists, one for the manager and one for the new hire, with an owner and a day number on every item. [PASTE DESCRIPTION]
2
Turn the task list into a first week schedule
Turn this list of onboarding tasks into an hour by hour schedule for the first three days, assuming a nine to five day with an hour for lunch. Leave two unscheduled blocks per day and tell me why each one is there. Flag anything that needs a person other than the manager to be available. [PASTE TASK LIST]
3
Write ninety day milestones a manager can actually measure
Write three ninety day milestones for a new [ROLE] that a manager can assess without a formal review process. Each milestone needs an observable outcome, a check-in date, and one sentence the manager can say if it is missed. No milestone may depend on something outside the new hire control.
4
Draft the welcome message and the team announcement
Write two short messages for a new [ROLE] starting on [DATE]. One goes to the new hire the week before they start and covers what to expect on day one, what to bring, and who will meet them. One goes to the team and introduces them in three sentences using only these facts: [FACTS]. No personal details beyond what I listed.
5
Find what breaks for a remote hire
This onboarding checklist was written for people in an office. Rewrite it for a fully remote hire in a different state. Flag every item that will silently fail remotely, such as equipment handover, document signing, and informal introductions, and give me a remote replacement for each. [PASTE CHECKLIST]

The remote rewrite prompt is the one I would run first. Checklists written for an office fail quietly rather than loudly: nobody reports that the introductions did not happen, they just do not happen.

Policy Draft Prompts

A model can write policy language. It cannot decide what your policy should be, and it should never be trusted on what the law requires. Describe the practice you already follow and ask it to turn that into a policy, with statutes explicitly off the table.

TaskPrompt to paste
Draft a policy from your actual practiceWrite a [TOPIC] policy for a US small business with [N] employees in [STATE]. Base it only on the practice I describe here: [DESCRIBE WHAT YOU ACTUALLY DO TODAY]. Short sentences, name the person responsible, say plainly what an employee should do. Mark anything that varies by state as [CHECK STATE LAW] rather than guessing, and do not cite statutes, regulations, or case law anywhere.
Rewrite a policy in plain EnglishRewrite this policy at an eighth grade reading level without changing its meaning. Keep every obligation, deadline, and exception exactly as it is. Then list any sentence where you are not certain that simplifying preserved the meaning. [PASTE POLICY]
Find the gaps in a handbookHere is the table of contents of our employee handbook. List the policies a US small business of our size commonly has that are missing here, sorted by how often the gap causes a real problem. Do not draft anything yet, and do not include policies that only apply to large employers. [PASTE CONTENTS]
Turn a policy into a manager one-pagerTurn this policy into a one page summary for managers. Cover what they must do, what they must not decide on their own, who to escalate to, and the three questions employees ask most about this topic. Keep it under 350 words. [PASTE POLICY]
Check two documents against each otherCompare these two documents and list every place where they contradict each other on the same subject. Quote both versions side by side and say which one is more specific. Do not resolve the contradictions for me. [PASTE BOTH]
The instruction that prevents most policy damage
Add do not cite statutes, regulations, or case law to every policy prompt. Invented or outdated legal citations are the most common serious failure in AI drafted HR documents, and a policy that quotes a rule incorrectly is worse than one that says nothing, because employees rely on it. Where a rule genuinely varies, have the model insert a bracketed placeholder and go check it against your state agency yourself.

The gap analysis prompt is the one I would run before writing anything. It is cheap, it takes one paste, and it surfaces the policies most small employers discover they need only when something has already gone wrong.

Performance Review Prompts

The problem a model solves in review season is phrasing, not judgment. You already know what you think. What is hard is saying it in language that is specific, observable, and does not read as a personality assessment.

1Turn rough notes into review language
PromptHere are my notes on an employee over the last six months, with the name removed. Turn each note into a review sentence that describes an observable behavior and its effect on the work. Remove personality judgments. If a note has no evidence behind it, tell me that instead of writing it up. [PASTE NOTES WITH NAMES REMOVED]The last instruction is the important one. It catches the impressions you have been carrying without examples.
2Soften the tone without losing the message
PromptRewrite this review paragraph so the tone is less harsh but the expectation is exactly as clear. If your rewrite makes what has to change any vaguer, say so explicitly and show me both versions. [PASTE PARAGRAPH]Softening almost always costs clarity. Making the model report the trade is how you keep it honest.
3Surface concerns you have not written down
PromptThis draft review is entirely positive and I have concerns I have not managed to put into words. Ask me five questions that will help me turn those concerns into specific, observable statements about work. Ask the questions one at a time and wait for my answer. Do not write any review text yet. [PASTE DRAFT]An all-praise review followed by a performance conversation three months later is the pattern this prevents.
4Fix vague goals
PromptTurn these three goals into specific goals, each with a measure, a deadline, and one named support step the manager commits to. Flag any goal whose outcome depends on something outside the employee control, and rewrite it so the part they control is what gets measured. [PASTE GOALS]Goals that depend on other teams are the main reason review conversations go sideways.
5Check the whole set for biased language
PromptReview these review paragraphs and flag any wording that describes personality rather than work, applies a different standard to different people, or uses words like abrasive, aggressive, emotional, bossy, or nice. Quote the phrase and give a behavior based replacement. Do not change anything you have not flagged. [PASTE PARAGRAPHS]Run this across all reviews at once rather than one at a time. The inconsistency only shows up in comparison.

Strip names before pasting review notes and use initials or a role label. The content of a review is exactly the kind of material that should stay inside systems you control, which is a point I come back to below.

Difficult Conversation Prompts

Use a model to rehearse, not to script. The value is in pressure testing what you plan to say and anticipating the responses, because the part that goes wrong in a hard conversation is almost never the opening.

1Draft the opening only
PromptI need to tell an employee that [SITUATION, NO NAMES, NO PERSONAL DETAIL]. Write an opening of no more than four sentences that states the issue, its effect on the work, and what has to change. Then list the three most likely responses and one line I could say to each. Do not script the rest of the conversation.Four sentences is deliberate. A long opening invites the other person to argue with the framing instead of the substance.
2Argue the other side
PromptHere is what I plan to say in a conversation with an employee. Respond as the employee would, taking the strongest reasonable position against me. Then, separately, point out anything I have asserted without evidence and anything that sounds like I have already made a decision I am pretending is open. [PASTE YOUR DRAFT]This is the single most useful prompt on the page. It finds the sentence that will cost you the conversation.
3Prepare a pattern conversation
PromptI need to raise a pattern rather than a single incident: [DESCRIBE THE PATTERN GENERICALLY]. Help me structure it so it does not turn into an argument about any one example. Give me an opening, a way to present three examples as evidence of a pattern, and a closing that sets a specific expectation with a review date.Have your three dated examples written down before you run this. If you cannot list three, it is not yet a pattern.
4Deliver a decision that is final
PromptI have decided [DECISION, NO NAMES]. Write four sentences that deliver it clearly without inviting negotiation. Then tell me the two questions I should expect and what a straight, non-defensive answer to each sounds like. Do not soften the decision and do not add reassurance I have not offered.Decisions delivered as though they are still open are the most common way a manageable conversation becomes a grievance.
5Write the summary afterward
PromptTurn these notes from a conversation into a short factual summary I can keep in the file: what was discussed, what was agreed, dates, and the next check-in. Facts only. No interpretation, no adjectives, no characterization of anyone mood or attitude. [PASTE NOTES WITH NAMES REMOVED]Write it the same day, put the name back in yourself, and store it with your other employment records.

Two boundaries. Never paste anything from a workplace investigation, including what somebody told you in confidence, and never ask a model to decide whether a complaint has merit. Rehearsal is a safe use; adjudication is not.

Companies Using FirstHR Onboard 3x Faster
Join hundreds of small businesses who transformed their new hire experience.
See It in Action

Survey Question Prompts

Survey writing is the task most improved by a second pair of eyes, and a model is a competent second pair of eyes for question design. It reliably catches leading questions, double-barreled questions, and questions nobody can act on.

1Write a short pulse survey
PromptWrite a ten question pulse survey for a [N] person US company with no HR department. Eight questions on a five point agreement scale and two open text. Cover workload, manager support, clarity of expectations, and whether people would recommend working here. No double-barreled questions, no leading wording, and no question that could identify a respondent in a company this size.The last constraint matters most below about thirty people, where a demographic question alone can identify someone.
2Fix a survey you already wrote
PromptReview these survey questions. Flag every leading question, every double-barreled question, every question with an unclear scale, and every question I could not act on if the answer were negative. Rewrite each flagged one and explain the problem in a single sentence. [PASTE QUESTIONS]The act-on-it test removes about a third of most first-draft surveys, which is the point.
3Build a thirty day new hire survey
PromptWrite eight questions to ask a new hire at day thirty. Four should test whether onboarding worked: access, clarity of the role, whether they know who to ask. Four should surface problems while they are still fixable. Keep the wording plain, avoid asking them to rate people by name, and make two of them open text.Day thirty is early enough to fix things and late enough that the answers mean something.
4Write exit interview questions
PromptWrite ten exit interview questions for a departing employee at a small US business. Order them from least to most sensitive. Avoid questions that invite a performance debate, avoid anything that reads as an attempt to talk them out of leaving, and include two that would surface a manager problem without naming anyone.Ask these in person if you can. The written version gets shorter answers.
5Turn responses into themes
PromptHere are anonymous open text responses from a staff survey, with any identifying details already removed. Group them into themes, count how many responses fall into each theme, and quote two representative comments per theme. Do not summarize away the negative ones, and tell me which themes are supported by fewer than three responses. [PASTE RESPONSES]The last instruction stops a single strongly worded comment being presented as a company wide theme.

Anonymity survives or fails at the paste. If a comment names a manager or describes a situation only one person could be in, remove it before the text goes anywhere.

What Never Goes Into a General Purpose AI Tool

Five categories of information should never be pasted into a general purpose chat window, regardless of how convenient it would be. None of them improves the quality of a draft, and all of them move sensitive data outside systems you control.

Employee and applicant identifiersSocial Security numbers, dates of birth, home addresses, bank details, driver license numbers, passport or immigration document numbers, and emergency contacts. None of it improves a draft, and all of it turns a writing tool into a place where identifying data now lives.
Medical and disability informationAccommodation requests, doctor notes, leave paperwork, test results, anything describing a health condition. Federal law requires employers to keep employee medical information confidential and separate from ordinary personnel records, and a chat window is neither.
Anything from a workplace investigationComplaints, witness statements, the names of the people involved, interim notes, the outcome. Investigation material is the single most sensitive category a small employer holds, and it is also the category people are most tempted to ask a model to summarize.
Salary data tied to named peopleIndividual pay, offer amounts, raise decisions, a spreadsheet with names beside numbers. If you want help with a pay structure, describe the bands and the roles with no names attached and no identifying detail.
Third party confidential informationCustomer lists, client contracts, anything covered by a nondisclosure agreement you signed. Your obligation to a customer does not pause because the paste was convenient, and most agreements say nothing about which tools are acceptable.
The working test: if you would not put it in an email to a supplier, do not put it in a general purpose chat window.

The reasoning is not only about training data. Consumer accounts may retain conversations and use them to improve the service unless you change the setting, and business tiers typically exclude inputs from training by default. Either way, the information has left your records and you cannot search it, produce it on request, or delete it with confidence.

Employee medical information carries a specific federal obligation. Employers must treat medical information obtained from any disability related inquiry or examination as a confidential record kept separate from ordinary personnel files, which is set out in the EEOC enforcement guidance on disability related inquiries. A chat log is not a confidential separate file, and it is not somewhere you can demonstrate access control.

The paste is the moment the risk is created
There is no undo. Once employee data is in a third party system it is out of your control, whatever the retention setting says. Build the rule into how people work rather than relying on judgment in a hurry: replace every name with an initial or a role label, delete identifiers before pasting, and never paste a document you did not write yourself. Write it down as a one paragraph AI use policy and give it to everyone who touches HR work, including managers who write their own reviews.

Practically, this means most HR uses of a general purpose model should be prospective rather than retrospective. Drafting a letter you are about to write is safe. Summarizing a file you already hold usually is not, because the file contains exactly the material that should stay in your own records.

A short written rule beats a training session. Mine is four lines: no names, no numbers that identify a person, nothing medical, nothing from an investigation. Anyone can hold that in their head at the moment they are about to paste, which is the only moment it matters.

Every Draft Needs a Human Reviewer

Everything a model produces is a first draft, and the fluency of the writing is not evidence that the content is right. Legal and policy content in particular has to be read by a person before it reaches an employee.

Three failure modes account for nearly all of the damage. Invented specifics, where the draft contains a number, a deadline, or a benefit you never supplied. Wrong law, where a statute is cited that does not say what the sentence claims or that no longer applies in your state. And borrowed voice, where the document reads like a company ten times your size, which employees notice immediately.

1
Read for specifics you never supplied
Every number, date, dollar figure, deadline, benefit, and job title. If you cannot point to where it came from, it came from the model. Delete it rather than editing it into something plausible.
2
Delete every legal citation
Whether or not you asked for one. If the document genuinely needs a legal reference, get it from the agency source yourself. Invented and outdated citations are the highest frequency serious error.
3
Check anything that varies by state
Leave, final pay, notice requirements, sick time, meal breaks, pay transparency. A model averages across jurisdictions. Your business exists in specific ones.
4
Rewrite one paragraph in your own words
Usually the opening. It reanchors the voice, and it forces you to actually read the draft rather than skim a document that already looks finished.
5
Route legal effect to a human reviewer
Offer letters, handbooks, discipline documents, separation language, accommodation responses, classification decisions. For a small employer this often means an hour with an employment attorney, which is cheap against the alternative.
6
Save the prompt, not just the output
The prompt is the reusable asset. Keep the ones that worked in a shared document so the next person does not start from a blank box and a vague request.

There is a harder line worth understanding, and it is the line between drafting and deciding. Using a model to write interview questions is drafting. Using it to score, rank, filter, or shortlist candidates makes it part of a selection procedure, and the federal Uniform Guidelines on Employee Selection Procedures treat any measure used as a basis for an employment decision as subject to adverse impact analysis. Nothing about that changes because the measure is software.

A growing number of states and cities regulate automated employment decision tools directly, and the pattern is consistent even where the details are not: notice to candidates, some form of bias testing, and a record you can produce. New York City requires employers using an automated employment decision tool to have it independently bias audited within the previous year, publish a summary of the results, and notify candidates in advance, under the city rules on automated employment decision tools. Illinois has amended its human rights statute to address AI in employment decisions and to require notice. Colorado passed a broad artificial intelligence act treating employment as a high risk use, then amended it and moved the effective date more than once, which is itself a useful signal about how settled this area is.

If you want a governance frame rather than a jurisdiction checklist, the NIST AI Risk Management Framework is voluntary, free, and organized around four functions: govern, map, measure, and manage. For a small business the useful takeaway is the first one. Decide in writing what these tools may and may not be used for before anyone starts using them, rather than after.

What worked for me
The habit that changed my output quality was writing the constraints before the request. I open every HR prompt with what the model may not do: no invented requirements, no statutes, no promises, no facts I have not supplied. It feels backwards, and it produces a noticeably duller first draft. It also produces a draft where everything in it came from me, which means editing is editing rather than fact checking a document that sounds authoritative and is quietly wrong in three places. The second habit is keeping a running document of prompts that worked, with the bracketed placeholders left in. Six months in, that file is more useful to my team than any of the outputs it produced.

Used this way, a general purpose model is a genuine advantage for a business without an HR department. It removes the blank page, which is the actual barrier, and it produces structure faster than any of us write it. What it does not do is take responsibility. That part is still yours.

Key Takeaways
Every prompt that works contains four parts: the role the model writes as, your business context, the constraints, and the output format.
Negative instructions do the heaviest lifting, because the default behavior of these tools is to fill gaps with confident invention rather than to ask.
For job descriptions, the trimming pass matters more than the drafting pass: ask which requirements are not truly needed in the first ninety days.
Offer letters are where invention is most expensive. Delete anything the model adds about tenure, promotion, or future pay rather than editing it.
Add do not cite statutes, regulations, or case law to every policy prompt, and have the model insert bracketed placeholders where a rule varies by state.
For difficult conversations, use a model to pressure test what you plan to say rather than to script the conversation.
Five categories never go into a general purpose chat window: employee identifiers, medical information, investigation material, named salary data, and third party confidential information.
Employee medical information must be kept confidential and separate from ordinary personnel records, which a chat log cannot satisfy.
Drafting and deciding are different. Scoring, ranking, or filtering candidates makes a tool part of a selection procedure subject to adverse impact analysis.
Every draft is a first draft. Check for specifics you never supplied, delete legal citations, verify anything that varies by state, and route legal effect to a person.

Frequently Asked Questions

What is the best ChatGPT prompt for HR tasks?

There is no single best prompt, but every prompt that works has the same four parts: a role, the context only you have, the constraints, and the output format. The role sets the register. The context is your headcount, industry, state, and what is already true in your business today. The constraints are the negative instructions that stop a model filling gaps confidently, such as do not invent requirements I did not give you and do not cite statutes. The format is the shape of the answer, such as six bullets, under 120 words, or a two column table with placeholders in square brackets. A prompt missing the context section produces a generic page. A prompt missing the constraints produces confident detail you never supplied.

Can I use ChatGPT to write job descriptions?

Yes, and it is one of the highest value uses for a small employer, provided you feed it your real duty list rather than asking it to imagine the role. The useful sequence is three prompts rather than one: draft from your rough notes, then ask the model to flag every requirement that is not genuinely needed in the first ninety days, then ask it to identify wording that could discourage qualified applicants. The second and third passes are where the value is, because inflated requirements shrink an applicant pool without improving hire quality. Review the pay range, the location, and any physical requirements yourself before the posting goes live.

Is it safe to put employee information into ChatGPT?

No, not into a general purpose consumer tool. Five categories should never be pasted in: employee and applicant identifiers such as Social Security numbers, dates of birth, home addresses and bank details; medical and disability information of any kind; anything from a workplace investigation, including names and witness statements; salary data tied to named people; and third party confidential information covered by an agreement you signed. Consumer accounts may retain conversations and use them to improve the service unless you change the settings, and even with retention off, the data has still left your systems. Strip names to initials or role labels, remove identifying details, and describe situations generically.

Can AI write an employee handbook or a company policy?

AI can produce a readable first draft of a policy, but it cannot decide what your policy should be, and it should not be trusted on the law. The reliable approach is to describe the practice you actually follow today and ask the model to turn it into policy language, with an explicit instruction to mark anything that varies by state as a placeholder rather than guessing. Tell it not to cite statutes at all, because invented or outdated citations are one of the most common failure modes. Anything with legal effect, including leave, pay, classification, discipline, and termination language, needs review by a person who knows your state law before it reaches employees.

Can I use AI to screen or rank job candidates?

Drafting and screening are different activities with different legal exposure, and the line matters. Using a model to write interview questions is drafting. Using it to score, rank, filter, or shortlist applicants makes it part of a selection procedure, and the federal Uniform Guidelines on Employee Selection Procedures treat any measure used as a basis for an employment decision as a selection procedure subject to adverse impact analysis. Several jurisdictions add their own rules on top. New York City requires an independent bias audit and advance notice to candidates for automated employment decision tools. Illinois has amended its human rights statute to address AI in employment decisions and require notice. Check your own state and city before automating any decision.

How do I stop ChatGPT inventing things in HR documents?

Give it negative instructions in the prompt and then verify the output anyway. The three instructions that do the most work are: use only the information I have given you, do not add requirements or terms I did not supply, and do not cite statutes, regulations, or case law. Asking a model to tell you when it is unsure also helps, for example telling it to list any sentence where simplifying may have changed the meaning. After that, read the draft specifically hunting for specifics you never supplied: numbers, deadlines, dollar figures, legal references, and named benefits. Every number and date in an HR document should be checkable against your own records before it goes out.

What HR tasks should I not use AI for at all?

Decisions, investigations, and anything where the answer depends on facts the model cannot see. Do not use it to decide who to hire, who to promote, who to discipline, or who to let go. Do not use it to analyze a complaint or summarize investigation material. Do not use it to determine whether an accommodation is reasonable, whether an employee is exempt from overtime, or how a state leave law applies to a specific person. Those are judgment calls with legal consequences, and a fluent wrong answer is more dangerous than no answer. Drafting, rewriting, structuring, and pressure testing your own thinking are the safe uses.

Do I need a paid or business AI account for HR work?

A business tier with administrative controls is a meaningful improvement over a personal consumer account, mainly because of data handling and settings you can enforce across a team rather than trusting each person to configure. Business and enterprise tiers typically exclude your inputs from model training by default and give an administrator visibility over accounts. That said, a paid account does not change what belongs in the tool. The five red line categories stay off limits regardless of tier, because the risk is not only training data, it is that identifying information about your employees now exists in a system you do not control and cannot search or purge on request.

Ready to transform your onboarding?

7-day free trial No credit card required
Start Your Free Trial