By McKinley Malbrough III, J.D., MS-HRM
Published November 6, 2026. Last updated November 9, 2026.
Short answer
Build a shared library where every prompt has a named purpose, an owner, a list of allowed and banned inputs, a required output review and a version date. Use AI for drafting and summarizing, never for deciding who is hired, disciplined or fired, and keep a person responsible for every result.
Build a shared library where every prompt has a named purpose, an owner, a list of allowed and banned inputs, a required output review and a version date. Use AI for drafting and summarizing, never for deciding who is hired, disciplined or fired, and keep a person responsible for every result.
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. This guide is for HR leads, owners and managers at 10 to 150 person organizations whose people are already using AI assistants for HR work, usually with no shared rules. It builds on my starter plan for AI in HR, which sets out what to hand to a tool and what to keep. This guide shows how to make the handoff repeatable.
Because individual use is invisible and inconsistent. One person writes a careful prompt and gets a useful draft. Another pastes an employee’s medical note into a chat window. A third asks the tool to rate a performance review. With no shared standard, the quality of the output depends on who happened to type, and the risks depend on what they happened to paste. A library solves three problems at once: it saves time by reusing what works, it improves quality by making prompts better over time and it controls risk by banning dangerous inputs in writing.
Use a five-part standard so every prompt has the same bones:
The fifth part matters most. A tool that lists its own assumptions gives your reviewer a place to start. It does not replace the reviewer, but it makes the review faster and more focused.
Ten tasks where drafting help is high and the decision stays with a person:
In each case the person checks the output against the facts, edits it and takes responsibility for it.
Write these into the library itself, so no one has to remember them:
California’s direction is toward human oversight. SB 947 states that an employer may not rely solely on an automated decision system to discipline or terminate a worker and requires human corroboration, effective July 1, 2027. The Civil Rights Council’s regulations on automated decision systems, in effect since October 1, 2025, apply discrimination rules to tools that make or support employment decisions. My guides on SB 947 and on AI hiring tools under the civil rights rules set out what they require.
Use one row or page per prompt with these fields: name, purpose, owner, version and date, the prompt text, allowed inputs, banned inputs, required review steps, and a note on what the output must never be used for. Store the library in a shared document with edit rights limited to two people. Everyone can read and copy; only the owners change. When someone improves a prompt, they send it to an owner, who tests it and updates the version.
Here is a prompt for the policy task, in my own wording. “You are helping an HR lead at a 40 person California employer draft a plain-language policy paragraph on remote work equipment. Use a ninth-grade reading level, under 150 words, neutral tone. Do not include names or specific dollar amounts. After the draft, list every assumption you made and any point that may need review under California law.” The prompt names the task, gives context without personal data, sets constraints, asks for a format and demands a review list. The reviewer then checks the draft against the actual policy and the law.
Use a short checklist on every use: Are the facts correct? Does it match our policy and California law? Does it contain anything that sounds like a decision about a person? Does the tone fit our company? Did I change anything that I should note? Record who reviewed it and when in a simple log: date, prompt name, reviewer, outcome. The log is cheap and it proves human review if a decision is ever questioned.
Five come up often. The library is too long, so people stop using it. It has no owner, so it drifts out of date. The red lines live in a separate policy nobody opens. The review step is optional, so it gets skipped under deadline pressure. Or the library is used as a shortcut around judgment, such as asking a tool to “rate this employee’s review” because it is faster. Keep the library short, assign two owners, put the red lines on the first page, make the review log part of the workflow and remind people that speed is never a reason to cross a red line.
It gives you a place to build in checks. For job descriptions, add a constraint that the tool flag exclusionary phrases and unnecessary requirements. For interview questions, require that each question tie to a stated skill and be asked of every candidate. For policy language, ask the tool to list who is most affected and what exceptions are needed. These do not remove bias, because a tool can carry bias from its training. They make bias easier for a human reviewer to see, and they keep the review focused on the questions that matter.
Hold one 45 minute session. Walk through the five-part standard, the red lines and two live examples, one that works and one that breaks a rule. Have each person run a prompt from the library on a task from their own week, using made-up data, and compare outputs. Then repeat a short refresher every quarter, since the tools change. Training should stress that the tool is a drafting assistant, not an authority.
Review the log monthly. Which prompts get used most? Which outputs needed heavy edits? Rewrite the weak prompts, retire the unused ones and add one new prompt each month for a task people keep asking about. Keep a short change history so you can see what altered the results. Check the tool provider’s data terms each quarter, because changes there can alter what you are allowed to paste.
The problem for an owner is that staff are already using AI tools on HR work, and nobody has decided what they may paste in or what must be reviewed by a person. A prompt library helps, but prompts only work if there is a policy behind them. The California employee handbook template ($149) is the editable policy base: the clauses that govern confidentiality, discipline and termination, written for California and annotated. Write your AI rules into that structure so the prompts and the policy say the same thing.
Write three entries using the five-part standard, one for a job description, one for a policy paragraph and one for an announcement. Add the red lines at the top of the document and share it with your team. If you want an outside review of how your people are already using AI and what rules to set, the HR compliance check-up includes it, and fractional HR support can build and maintain the library.
Related reading: a green, yellow and red map of where AI helps HR; five reusable AI skills with refusal rules.
Related reading: The HR AI Field Guide, my book of plain-English terms, prompts and safe rules for AI in HR.
For a related angle, see writing one AI assistant spec you can port between tools.
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.
Share this article
This article is general information, not legal advice. Laws change and every situation is different; for advice on yours, talk with an employment attorney.
Related questions
November 6, 2026
Generative AI can draft, summarize and organize HR work, and it can also leak data and repeat bias. A starter plan: five tasks to hand over, five to keep and a one-page…
November 4, 2026
SB 947 takes effect July 1, 2027. If software drove your discipline or termination, here is what the employer must do, what you can request in writing, what to save and who…
October 26, 2026
Most small California employers fall outside the privacy agency's new AI rules. They do not fall outside the Civil Rights Council's. What to check this month.
