13 AI prompts for HR: Organized by task and ready to run

AI has already found its way into HR’s day-to-day work. In a 2025 General Assembly survey of 288 HR professionals in the U.S. and U.K., 82% said they used AI at work. The same research found that only 30% had received comprehensive, job-specific AI training.

That gap matters. Getting useful results requires more than typing a request into a blank chat. You need to give the tool enough context, know which information should stay out of the prompt, and review the output with the same judgment you would bring to any other HR work.

This library gives you 13 prompts for recruiting, employee communications, policy, performance management, employee relations, people analytics, and a custom benefits assistant. Each one is designed to be adapted—not pasted blindly—and every consequential output stays subject to human review.

Best practices for prompting AI in HR

Build every prompt with CRAFT: context, role, action, format, and tone

“Vague prompts, vague outputs, and they might erode trust.” That is how Jenelle Buatti, an instructor for our AI for HR Professionals course, describes the stakes.

A weak prompt for a marketing tagline may waste ten minutes. A weak prompt for a policy announcement could confuse an entire workforce. CRAFT gives the model the direction it needs while making your assumptions easier to inspect.

Context is the part people most often skip. Treat each new chat like a new teammate: Provide the audience, background, constraints, and definition of a useful result.

Consider a performance improvement plan (PIP). A risky prompt might include the employee’s name, exact results, dates, and manager comments. The following version can produce a useful structure without those details.

Act as a manager drafting a performance plan for a sales director. The employee has missed the quarterly sales quota for two consecutive quarters. Draft a section focused on pipeline management and prospecting activity without referencing any names, exact financial figures, or precise dates. Separate observed performance from assumptions about its cause.

The prompt supplies a role, relevant context, a specific action, and clear constraints. Before running it, however, you still need to know whether the tool is approved for the information involved.

[[cta]]

Red light or yellow light: What HR data can go into an AI tool?

Use the course’s red-light/yellow-light model as a conservative starting point:

• Red light: A consumer or public AI workspace that your organization has not approved. Do not enter confidential company information, personal data, personnel records, medical information, investigation details, or other sensitive material.
• Yellow light: An organization-approved business or enterprise workspace. Confirm the tool’s settings, contract, retention controls, access permissions, and your internal policy before entering company data. Approval is not permission to paste anything.

AI products handle data differently. For example, conversations in a personal ChatGPT workspace may be used to improve OpenAI’s models unless the user opts out. Follow your organization’s approved tool and data-handling policies before using HR information.

Chiara Barden, another instructor in the course, identifies five categories that HR professionals should remove or generalize before working with AI:

Free-text fields deserve a second look. An export may have clean structured columns while comments still contain names, locations, health information, or identifying anecdotes. Anonymization is not simply deleting the “name” column.

AI prompts for HR, organized by task

1. Recruiting prompts for job descriptions, screening criteria, and interview rubrics

Drafting a job description and ranking candidates may feel like similar forms of “recruiting help,” but the risks differ. One produces language for a person to approve. The other can materially affect who receives an employment opportunity.

Start with drafting and structured interview preparation:

Act as a talent partner hiring a mid-level [function] professional at a [industry] company. Context: The role reports to [title]; the required skills are [three to four skills]; the work location is [location or remote]; and the compensation band is fixed by internal equity. Action: Draft a posting with a role summary, responsibilities, must-haves, and nice-to-haves. Then flag any requirement in your draft that could unnecessarily screen out a qualified candidate. Do not invent responsibilities, benefits, or company details. Format: Job posting followed by a review checklist. Tone: Plain and specific, with no superlatives.

Act as a hiring manager building a structured interview rubric from the job posting below. Context: The process has three rounds, and interviewers need to assess the same job-related criteria. Action: Identify five criteria drawn only from the posting’s stated requirements. For each one, create behavioral anchors showing what a one, three, and five response might demonstrate. Mark any criterion that cannot be supported by the posting. Format: Table. Posting: [Paste approved posting].

AI can help draft the instrument used to evaluate candidates. A person should validate the criteria, apply the rubric, read the source materials, and make the decision.

In New York City, Local Law 144 restricts the use of covered automated employment decision tools unless a recent bias audit has been completed, the required information is public, and notices have been provided. City guidance specifies a 10-business-day notice before use. Coverage depends on the tool and circumstances, so confirm the law’s application before deploying one.

Federal anti-discrimination laws also apply when technology is used in hiring. The EEOC and Department of Justice warn that algorithmic tools can screen out people with disabilities and that employers may need to provide reasonable accommodations or alternative evaluation methods.

2. Employee communication prompts that sound like you

The fastest way to improve AI-generated communications is to give the model approved examples of your actual voice. “Warm but professional” is abstract; two or three messages you wrote give the model something concrete to match.

Maisha L. Cannon, an instructor for the course, teaches a lighter framework for content drafting: POV, which stands for perspective, objective, and voice.

• Perspective: Who is speaking, and whose experience should shape the message?
• Objective: What should the reader understand, feel, or do?
• Voice: How should the message sound?

Start with an onboarding note:

Perspective: A warm, organized manager writing to someone who starts Monday. Objective: Help the new hire feel expected and prepared. Voice: Warm and concise, with no corporate filler. Write a welcome email to a new [role] starting [general date or approved date]. Include the first-day schedule ([details]), who will greet them ([role or approved name]), what to bring, and a one-sentence explanation of why the team is glad they are joining. Keep it under 200 words and match the approved examples below. [Paste two examples that contain no confidential or unnecessary personal information.]

The same structure can support a 30-60-90-day plan or buddy-program introduction. For an announcement that needs more control, return to CRAFT:

Act as a benefits communications specialist. Context: Open enrollment runs [dates]; the medical carrier is unchanged; the [plan] deductible changes to [amount]; and [new benefit] is being added. Employees must take action even if they do not want to make changes. Action: Draft the announcement, then add a three-bullet summary at the top. Format: Email with a subject line. Tone: Clear and upbeat, with no hype. Do not add plan details that are not provided, and flag any information an employee would still need.

Act as an HR business partner writing to people managers. Context: [Policy] changes on [date]; the business reason is [reason]; and managers may hear [likely objection]. Action: Draft a short memo explaining what changed and why, then write three talking points a manager can use in a one-on-one conversation with a skeptical employee. Format: Memo followed by talking points. Tone: Direct, supportive, and non-defensive. Do not make legal promises or invent exceptions.

Before sending any AI-assisted communication, ask three questions: 

• Does it sound like us? 
• Does it answer the first question employees will have? 
• Does it name the right place to get help?

3. Policy and handbook prompts that pressure-test the draft

Asking AI to “make this policy announcement better” often produces surface-level reassurance. A critical persona adds tension to the review by asking the model to articulate a skeptical reader’s objection before suggesting a revision.

Buatti demonstrates the approach with a "use-it-or-lose-it" paid time off announcement. Here is a polished version you can adapt:

Act as a skeptical employee who has experienced a use-it-or-lose-it PTO policy and views it as more restrictive than a carryover policy. Read the announcement below. First, explain why an employee might view the change negatively, quoting the wording that creates that impression. Then suggest revisions that acknowledge the concern without minimizing it or making promises the policy does not support. Announcement: [Paste approved, nonconfidential draft].

The role is adversarial rather than agreeable. The action asks for objections before revisions, and the context gives the model a specific concern instead of generic dissatisfaction.

For a first policy draft:

Act as an HR policy writer preparing a first draft of a [policy topic] policy for employees in [jurisdictions]. Context: Current practice is [practice], the change is driven by [reason], and the organization believes [named statute or rule] may apply. Action: Draft the policy section using only the supplied facts. Then list every legal, jurisdictional, operational, or employee-relations point that requires confirmation. Do not invent citations or state that the draft is compliant. Format: Draft policy followed by a flagged-items checklist. Tone: Formal and plain.

AI can help identify questions and organize a first draft. It cannot verify that a policy complies with every applicable law. Route the result to the HR, legal, or compliance owner responsible for approval, and verify every citation independently.

4. Performance management prompts for reviews, feedback, and coaching

The strongest performance prompt is not the one with the best adjectives. It is the one that defines the evidence, removes unnecessary identifying details, and tells the model what it must not infer.

Act as an experienced people manager. Context: A [level] individual contributor has [observable behavior, such as missed three sprint commitments without flagging risk early]; the business impact is [impact]; and the issue has been discussed previously. Do not include names, exact dates, medical information, or speculation about motivation. Action: Draft feedback that separates observed behavior from interpretation and proposes two concrete changes with a check-in point. Format: Short summary followed by three talking points for a live conversation. Tone: Direct and supportive.

The phrase “separates observed behavior from interpretation” helps keep unsupported judgments—such as “seems disengaged”—out of a document that may become part of an employment record.

For aggregated feedback:

Act as an HR business partner. Context: Below are [number] approved, anonymized feedback inputs about a [role-level] population. Identifiers, exact dates, medical information, and compensation data have been removed. Action: Summarize recurring themes and quote the language supporting each one. Do not infer causes, rank individuals, or draw conclusions that the inputs do not support. Label any theme appearing in fewer than three inputs as limited evidence. Format: Table with theme, supporting evidence, and confidence. Tone: Neutral.

Read every free-text field before pasting. Then check the output against the originals. The version that enters an employee record must reflect the reviewer’s judgment and the underlying evidence—not an AI-generated inference.

5. Employee relations prompts for questions that belong with counsel

In employee relations, a useful prompt often has a deliberately narrow job: Prepare questions for a qualified professional without asking AI to decide the case.

Act as a research assistant helping an HR professional prepare for a consultation with licensed employment counsel in [jurisdiction]. Context: The employer is considering [category of action] involving an employee at [role level] with [broad tenure range]. No names, dates, health information, investigation details, or identifying facts are provided, and none should be assumed. Action: List the factual and legal questions to raise with counsel before proceeding. Group them under process, documentation, notice and pay, employee rights, and risk. Do not advise whether to proceed, predict an outcome, cite a law you cannot verify, or draft an employee-facing document.

Keep medical and accommodation information, active investigations with identifiable participants, and disciplinary decisions out of unapproved AI tools. AI prepares the questions; a qualified person answers them and owns the decision.

6. People analytics prompts for turnover, engagement, and headcount reporting

A dataset can have no names and still identify people. Small groups, exact dates, rare job titles, locations, and free-text comments can make an individual recognizable when combined.

Prepare the export before writing the prompt. Remove direct identifiers, generalize dates and compensation, inspect comments, and combine groups that are too small under your organization’s privacy standard.

Act as a people analytics partner. Context: The approved dataset covers voluntary exits from [department group] during [period], aggregated by tenure band, level, and manager span. Direct identifiers, exact dates, individual compensation, and free-text comments have been removed, and groups smaller than [organization-approved threshold] have been combined. Action: Identify patterns in the supplied data. For each one, explain whether the evidence appears strong enough to investigate or should be treated as preliminary, and list the additional data needed to test the leading explanation. Do not speculate about individual motivations or protected characteristics. Format: Table with pattern, evidence strength, limitations, and next check. Tone: Neutral and analytical.

AI may help surface a pattern. It cannot determine whether the pattern is fair, what caused it, or what leadership should do about it without business context and accountable human judgment.

Increase consistency with a custom GPT: Where to start

The same CRAFT elements that improve one prompt can configure an assistant for a recurring, tightly scoped task. Maggie Wong, an instructor for the course, uses a benefits companion as the worked example.

OpenAI’s current GPT creation instructions describe custom GPTs as combinations of instructions, knowledge, and capabilities. Access depends on the user’s plan and workspace permissions.

1. Define the use case. Limit the assistant to a narrow set of approved benefits questions.
2. Write the name and description. Make the scope obvious to an employee scanning a list of tools.
3. Add instructions. State what the assistant may answer, what it must not do, which sources it may use, and when it must escalate to HR.
4. Add approved knowledge. Upload only current documents that your organization has approved for the workspace and intended audience.
5. Review capabilities. Disable capabilities the use case does not require, including web search if answers must come only from approved plan documents.
6. Add conversation starters. Give employees two or three safe examples that demonstrate the assistant’s intended scope.
7. Test before sharing. Include normal questions, ambiguous questions, requests for personal recommendations, outdated information, and attempts to move beyond scope.

Adapt this instruction block:

You are a benefits information assistant for [Company] employees. Answer questions about our health, dental, vision, and retirement benefits using only the approved documents in your knowledge base. Name the document and section supporting each answer. Do not give legal or tax advice, recommend a plan for an individual, interpret medical facts, or invent a coverage detail, amount, deadline, or exception. Do not request names, medical information, salary information, dependent details, account numbers, or other personal data. Remind users not to include personal information in their questions. If the approved documents do not answer the question, say so and direct the employee to [HR contact channel]. Keep answers under 150 words in plain language. When a deadline is involved, state the exact date from the source document.

Maintenance is part of the product. Test representative employee questions regularly, review instructions and knowledge files each quarter, conduct a fuller annual audit, and replace documents as soon as a plan changes. Assign an owner before the assistant is published.

FAQ: Common questions about using AI prompts in HR

Will using AI for HR tasks get me in legal trouble?

Using AI is not automatically unlawful, but the way it is configured and used can create privacy, discrimination, employment, records-management, and contractual risk. Use approved tools, minimize the data you provide, keep meaningful human review over employment decisions, and obtain legal advice for consequential or regulated uses.

What HR data is safe to paste into ChatGPT?

There is no universal list because the answer depends on the workspace, settings, contract, employer policy, jurisdiction, and purpose. In a personal or unapproved workspace, do not paste confidential HR or employee information. Even in an approved business workspace, use only the minimum data authorized for the task.

Is AI going to replace my HR job?

AI is more likely to change the mix of HR work than eliminate the need for human judgment. In General Assembly’s 2025 survey, 69% of respondents said AI freed time for strategic work, and 41% reported working fewer hours. The opportunity is to move routine production work toward drafting and automation while strengthening the skills AI cannot own: judgment, trust, context, and accountability.

Why does AI-generated HR copy sound generic, and how do I fix it?

Generic prompts produce generic language. Supply the audience, purpose, constraints, and two or three approved examples of your writing. Then ask the model to identify the voice characteristics it is following before drafting. A human should still revise the result for accuracy, empathy, and the circumstances behind the message.

Do I need a business AI workspace for HR work?

Use the environment your organization has approved for the information and task involved. Business products may provide stronger default privacy, security, administration, and retention controls, but the plan name alone does not determine whether a use is appropriate. Confirm the configuration, permissions, contract, and internal policy before using HR data.

If you want to build these skills into a repeatable workflow, the AI for HR Professionals course covers prompting, data safeguards, practical use cases, and human oversight across the HR function.