Keep the file out of the chatbot
Someone on your team will paste an employee file into a chatbot this week, because summarizing it by hand felt slow. Your job this hour is not a coding puzzle and not a quiz about loops. It is the HR outcome: draft faster without putting someone's private file in a prompt. The check that stops that paste is what you build here.
You're in the people studio. What you're doing this lesson: install one rule before any recruiting workflow — no personal employee data in prompts — and see why a rule held in memory fails on the busiest Thursday. You do not need HRIS software or a coding background to start.
Your function is adopting AI faster than any survey predicted. SHRM's 2025 Talent Trends survey — 2,040 HR professionals — put AI use at 43% of organizations, up from 26% in a single year, with recruiting leading: 66% of AI-using recruiters draft job descriptions with it, 44% screen resumes with it, and 89% say it saves time. You've probably already used it this week, for the JD you didn't want to write from the last posting's corpse.
Here's what makes HR different from every other desk in the company doing the same thing: your inbox is where the protected data lives. PII (personally identifiable information — a name tied to a medical note, a date of birth, an investigation file). An accommodation request with a diagnosis attached. Investigation notes from the thing in the warehouse. A background-check report the FCRA has opinions about. The candidate file with a date of birth in it. Marketing can paste their whole campaign brief into a chatbot and the worst case is embarrassment. Your worst case has a statute number.
So the people function runs on one rule that outranks every workflow in this chapter:
No personal employee data in prompts. Not names tied to issues, not medical information, not employee-relations material. Ever.
The problem with that rule is where teams keep it: in memory. It gets announced in a team meeting, lands in a slide, and holds beautifully until the Thursday when three reqs are open, open enrollment starts Monday, and someone pastes an FMLA letter into a chatbot because summarizing it by hand felt slow. Rules held in willpower fail exactly when the workload is heaviest — which in HR is always.
Rules held in a check don't have Thursdays. A gate reads the payload before anything is sent, checks it against a never-list, and returns BLOCK with a reason. It isn't smarter than you. It's just never tired, never behind on reqs, and never sure that "just this once" is fine.
This lesson builds that gate. First the firewall as a list — what may enter a prompt, what never does — then the scanner, then the failure mode where a gate checks some categories and quietly waves the rest through. By the checkpoint you'll have a prompt_gate() that audits a batch of real-shaped payloads and tells you what it stopped.
One category gets special handling: employee relations takes nothing, not even a summarization pass. That's step six, and it has no exceptions on purpose.