The investigation file has your name on it. Not the AI's.
Thirty compliance framed prompts, seven templates and three checklists for running employee relations cases with an AI assistant, built so the documentation holds up for whoever reads it after you.
Written by someone who spent about fifteen years working in HR inside large companies, several of them Fortune 100, and who now builds AI systems for a living.
Generic prompts give you generic language
In an employee relations file, generic language is the whole problem.
"The respondent was unprofessional" is an opinion wearing a finding's clothes. It will not survive being read back to you by someone with a reason to take it apart. And you are the one who signed it.
That is the part nobody selling AI productivity mentions. The tool does not sit in the appeal hearing. You do. When someone asks how a paragraph got written and what you did to check it, the answer has to be better than the truth most people currently have, which is that they asked a chatbot and it sounded right.
The deadline already passed
On 2 August 2026 the EU AI Act became applicable. It treats AI used to make decisions about promotion, termination and the monitoring of performance as high risk, and it puts named duties on the employer deploying it. Human oversight by someone with the competence and authority to exercise it. Logs kept. Workers told.
If your organisation has people in the EU, that is live now. If it does not, the discipline still holds, because the question an appeal body asks has never been jurisdictional. What did you do to check it.
Here are the sources. Read them yourself rather than taking my word for it.
- Annex III, point 4(b) classifies as high risk "AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics or to monitor and evaluate the performance and behaviour of persons in such relationships." Annex III
- Article 113: "It shall apply from 2 August 2026." Article 113
- Article 26(2): "Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support." Article 26(6) requires logs kept for at least six months. Article 26(7) requires workers and their representatives to be informed before use. Article 26
- Article 99(4) sets fines for breaching those deployer duties at "up to 15 000 000 EUR or, if the offender is an undertaking, up to 3 % of its total worldwide annual turnover for the preceding financial year, whichever is higher." Article 99
This pack is not legal advice and it is not a compliance certification. Whether a specific tool in your organisation falls inside that classification is a question for your counsel.
A working system, not a cheat sheet
Every prompt carries two annotations. One tells you why the wording is what it is and why the obvious alternative is worse. The other tells you what breaks if you run it raw.
Prompt 18 of 30, free
This is the one that pulls a draft finding apart into fact, inference and opinion. Run it on something you have already written and filed. You will find something.
You are assisting a qualified HR investigator. Separate fact from inference from opinion in draft findings. You do not decide whether anything is proved. INPUT Draft findings text: [PASTE] Evidence references available: [LIST] TASK Take every statement in the input and classify it into exactly one of four categories. 1. FACT SUPPORTED BY EVIDENCE. A statement about what happened that can be traced to a named evidence reference. Rewrite it in the form: [WHAT HAPPENED], on [DATE], evidenced by [REFERENCE]. 2. FACT ASSERTED WITHOUT EVIDENCE. Reads as fact, traces to nothing in the list. Rewrite it as an open question and name the evidence that would settle it. 3. INFERENCE. A conclusion drawn from facts. Rewrite it to show its structure: because [FACT, REFERENCE] and [FACT, REFERENCE], it is more likely than not that [INFERENCE]. Then state what alternative explanation was considered and why it is less likely. 4. OPINION OR CHARACTERISATION. A judgment about a person, their motives, their attitude or their character. Either delete it, or rewrite it as the observable conduct underneath it plus its evidence reference. FLAG SEPARATELY - Every adjective doing evidentiary work: aggressive, hostile, unprofessional, inappropriate, dismissive, erratic. Replace each with the conduct described. - Every statement about what a person knew, intended, meant or felt, unless they said so and it is referenced. - Every use of "clearly", "obviously", "evidently", "of course". These signal an unstated inference. - Every passive construction that hides who did what. - Every generalisation across incidents that is not itself supported. OUTPUT The four classified lists with rewrites, the flag list, and a clean redraft containing only category 1 and category 3, with every category 3 item showing its reasoning.
Use role labels, not names. The pack explains why on page one.
Someone who has had to defend a file
I am Boubacar. I spent about fifteen years working in HR inside large companies, several of them Fortune 100. Investigations, performance cases, terminations, and every piece of documentation that comes attached to those.
Now I build AI systems, from Salt Lake City.
That pairing is the reason this exists. Most people writing AI prompt packs have never had to sit across from someone and explain a finding. Most people who have done that are not building with these tools daily. This pack lives where those two things overlap, and there are not many of us standing there.
Do not buy this if
- You want faster memos. A nine dollar general prompt pack will do that, and it will do it well. This one is slower on purpose. It makes you write down what you checked.
- You want a compliance certificate. There is no such thing here. This is a documentation discipline, and the pack says so on every page.
- You want something that decides for you. Several of these prompts are written to refuse. The credibility framework will not tell you who to believe. That is deliberate, and if it frustrates you, this is not your pack.
$19 for the first fifty
Then it goes to $29. No subscription, no seat count, no login. You get the pack and it is yours.
- Everything is written to be edited. Your policy language, your jurisdiction, whatever your counsel prefers.
- Works with the assistant your employer already approved. Plain text, paste it in.
- Consultants may use it inside their own client engagements. Not for resale or redistribution.
Questions people actually ask
No, and the pack says so on every page. Run anything consequential past your employment counsel. What this changes is that when you do, there is something coherent in front of them instead of a folder and a story.
Yes. Every prompt is plain text with bracketed inputs you fill in. Paste it into whatever your organisation has approved. The pack also sets out how to use role labels so you are not putting names into a tool that should not hold them.
Only if your organisation has people or operations in the EU, and more of them do than realise it. The documentation discipline is the same either way. Nothing in the pack is written for one jurisdiction, and the parts that touch law are flagged for your counsel rather than answered for you.
One page you open in a browser. Contents down the side, everything collapsible, a copy button on every prompt and every template. It works on a phone, which matters more than it should, because a lot of this gets read on the way to a meeting.
Because I want the first fifty people using it and telling me where it is wrong. The price goes to $29 after that, and it will not come back down. If you want the honest version: early buyers are worth more to me as readers than as revenue.
The pack ends with a list of the cases that go outside what any template can carry, and a link to book a working session with me. Bring the case, including the parts that are not tidy. Anonymised facts only.
One file, done properly, pays for this fifty times over
The next case is already coming. You can run it the way you ran the last one, or you can run it with the questions already written down.
Get the pack, $19 First fifty buyers. Then $29.