Five Reusable AI Skills Every Small HR Team Can Build, With Refusal Rules

A reusable AI skill is a written set of instructions for one task, with a purpose, an input list, an output format, a refusal rule and an owner. Build five first: job posts, policy questions, review drafts, exit summaries and meeting notes. The refusal rule matters most, because it tells the tool to say it does not know instead of guessing.

I am Mac Malbrough. I hold a J.D. and an MS-HRM, I am not an attorney, and this is general information, not legal advice. I use saved instruction sets like these in my own work, and I wrote this for small HR teams who retype the same long instruction every week and get a different result each time. A skill fixes that. It makes the instruction the thing you manage, instead of the answer.

The problem this solves: Your team retypes the same long AI instruction every week, gets a different answer each time, and a policy chatbot with no refusal rule will guess about leave and pay.

How UHR solves it: The AI Job Posting Prompts ($9) and AI Meeting Recap Prompts ($7) give you tested wording for your first two skills, and the California employee handbook template ($149) gives the policy skill an accurate source to quote.

What is a reusable AI skill?

It is a saved set of instructions that tells an AI tool how to do one specific task the same way every time. Think of it as a job aid for a very fast, very literal assistant. You write it once, test it, give it an owner and put a date on it. Then anyone on the team can use it and get a consistent result.

A skill is different from a prompt library. A library is a collection of requests. A skill is a task with a standard attached. If the prompt asks for a job post, the skill says what the post must contain, what the tool must never invent, and what to do when information is missing. The ten prompts in a typical library become five skills when you decide what each task must always do.

For the broader library approach, including the red lines California puts around people decisions, see a safe prompt library for HR tasks.

Terms used here, such as custom assistant, refusal rule and review step, are defined in the Urban Glossary.

What parts does every skill need?

Every skill, however simple, needs seven parts. If one is missing, the result will vary.

1. Purpose. One sentence on the task and who uses the result.
2. Allowed inputs. What the user may provide, and what they may not, such as medical details, Social Security numbers or complaint facts.
3. Source rule. Where answers may come from, such as only the pasted policy.
4. Output format. The headings and length you expect.
5. Refusal rule. What the tool does when it lacks information or is asked something outside the purpose.
6. Review step. Who reads the result before anyone sees it.
7. Owner and date. A named person and the date of the last review.

What does the policy question skill look like?

This is the skill that most needs a refusal rule, so it is the best place to start learning. The task is to answer an employee’s question using only the policies you paste into the tool.

The purpose is to answer routine first-line questions. The allowed input is the employee’s question and your handbook text. The source rule is that every answer must come from the pasted text and must quote the section it used. The refusal rule is the key: if the pasted policies do not answer the question, the skill says that it cannot answer and routes the question to HR. It must not guess, fill gaps from general knowledge or describe what the law requires. The review step is a weekly read of every question the skill could not answer, which also shows you where your handbook has gaps.

That refusal rule keeps a tool from telling an employee that they are entitled to something your policy does not give, or that they are not entitled to something the law does give. California leave, pay and break rules are exactly the areas where a confident wrong answer causes harm.

What does the job post skill look like?

The task is to turn a hiring manager’s intake notes into a clear, fair posting. The allowed inputs are the duties, required qualifications, work location and schedule. The source rule is that the tool may only use what the manager provided. The output format has headings for summary, duties, requirements and preferred skills.

The refusal rule is that the tool must not invent qualifications, years of experience or benefits, and must ask for the pay scale if it is missing. California requires pay scale information for postings by employers of a certain size, covered in Labor Code Section 432.3, so a skill that leaves the field blank should say so out loud. The review step is a person who confirms the duties and requirements and adds the final pay range.

What does the performance review draft skill look like?

The task is to turn a manager’s own notes into a structured draft with strengths, gaps and next steps. The allowed input is the manager’s observations about work, with dates. The source rule is that the draft may only use those notes. The refusal rule is that the tool must not add observations, rate the employee or use language about age, health, family status or other protected traits.

The review step belongs to the manager, who edits and owns the final text. This is a yellow task and not a red one only because the manager makes the judgment. The skill organizes the words.

What does the exit summary skill look like?

The task is to turn offboarding information into a structured summary for the HR file: last day, final pay items to check, equipment to return and themes from the exit conversation. The allowed inputs are your own checklist answers and notes. The refusal rule is that the tool must not state what final pay is owed or decide whether any issue needs escalation. It may flag words that suggest a complaint or a legal concern, so a person reads the notes.

The review step is an HR person comparing the summary to the final pay rules. California has strict timing for final pay, and the summary is not a substitute for checking it.

What does the meeting notes skill look like?

The task is to turn your own notes into a short summary, a list of decisions, owner-tagged action items and unresolved questions. The refusal rule is to mark anything it is unsure about as unresolved rather than choose an owner or a date. The allowed inputs exclude meetings where confidential employee matters were discussed. The review step is the meeting owner reading the actions before they are sent.

For the privacy side of note-taking tools, see AI note-takers in HR meetings.

How do you test and retire a skill?

Test every skill on five real examples before anyone relies on it. For each, check four things: did it use only the allowed sources, did it follow the format, did it refuse when it should have and did the reviewer find anything that would have caused harm if it had gone out?

Try to break the refusal rule on purpose. Ask the policy skill a question your handbook does not answer. If it answers anyway, the skill is not ready.

Review every skill on a schedule, at least quarterly, and whenever your policy or the law changes. Retire a skill when the task changes or when the tool no longer follows the instructions. A skill that is out of date is worse than none, because people trust it.

How does the UHR handbook template give the policy skill something to stand on?

The problem for most small teams is that a policy question skill is only as good as the policy it reads. If your handbook is out of date or silent on California rules, the skill will refuse often or, if it is badly written, answer from guesswork. The California employee handbook template ($149) gives you an annotated, editable policy base written for California, so the source rule has something accurate to quote. The weekly review of refused questions then becomes your handbook improvement list.

For the skills themselves, AI Job Posting Prompts: 12 Prompts for Clear, Fair Job Descriptions is $9 and AI Meeting Recap and Action Item Prompts: 10 Prompts for Facilitators and Managers is $7. Both give you the prompt text to turn into the first two skills, and the free AI Prompt Pack Sampler: 12 Free Claude Prompts is a safe place to practice.

What should you do this week?

1. Pick the one task your team repeats most and write its seven parts on one page.
2. Write the refusal rule first, in plain words, before the rest of the instructions.
3. Test it on five real examples, including one it should refuse.
4. Name an owner and a review date, and share it with the team.

If you would like help turning your most repeated HR tasks into skills with proper review, the HR compliance check-up can show you where to start, and fractional HR support can build and maintain them with you.

For a related angle, see writing one AI assistant spec you can port between tools.

Sources

• California Labor Code Section 432.3, read on the California Legislative Information site
• California Civil Rights Department, automated decision system regulations

McKinley holds a J.D. but is not a licensed attorney. Articles here are general information, not legal advice. For your specific situation, talk with an employment attorney.

Data current as of November 2026. Sources are linked where each figure appears.