Lesson 5.1: Traffic light in practice

Goal: in 10 seconds know whether a text belongs in a public chat.

We've used the traffic light since module 1. Now we make it permanent:

Lesson 5.2: What if a doctor still drops a patient case into a prompt?

Goal: don't secretly feel guilty, do react practically.

It happens: a doctor emails you a text containing a patient name, age, diagnosis and the question “Can you have ChatGPT rewrite this?” You are then at the choice point. The rule of thumb:

1. Pause. Don't paste immediately “because it's urgent”. Ask the doctor whether it can go anonymous.
2. Anonymise. Replace name, BSN, date of birth, region and specific diagnosis combinations.
3. Check policy. Unknown? A short question to IT / privacy officer about which route is approved in your hospital.
4. Document. Keep a brief record of what you asked and what answer you got. If a privacy question comes up later, it's on paper.

Lesson 5.3: Mistral as the EU route

Goal: know where “the European variant” fits in your vocabulary, even if you don't use it yourself yet.

Mistral is a French/EU AI company that makes the “Le Chat” chat product and accompanying models. For healthcare institutions with a European preference (data residency, GDPR as starting point rather than later sticking point) Mistral is often on the shortlist. For you as a staff secretary it's important to know the name, IT/procurement may have questions about it.

Lesson 5.4: Build your own prompt library

Goal: 10–15 templates that carry you through a work year.

A prompt library isn't a secret art. It's a work document (Word, OneNote, Notion, or inside a Claude Project) with your best prompts. Without a library you start over every time; with a library you pluck the right prompt in 30 seconds.

Suggested structure:

1. Mail (5 prompts): short internal email, polite refusal, onboarding email new staff member, email on behalf of department head / training supervisor, external email.
2. Long text (3 prompts): minutes from transcript, agenda from mail thread, guideline on one A4.
3. Excel/Office (3 prompts): invent table + formula, rewrite slide, document with track changes.
4. Quick help (2-4 prompts): tone shift, summary of a conversation, brainstorm questions for a meeting.

Lesson 5.5: Automation bias: the creeping risk

Goal: know why after three weeks of AI use you read less critically, and how to counter that.

An unexpected risk most people underestimate: automation bias. The more often you see AI output that's good enough, the more you tend to accept it without conscious checking. “It said so, so it must be right.” This creeps in. A week after the course you're still critical, after three months less so, after a year you skim an email and forward it.

Research on healthcare professionals shows this really happens: the first time an AI gets something wrong it's noticed, by the tenth time it isn't. For your work this means you need a few light structural habits:

• Reading pause. Between “AI done” and “you send” there's at least 30 seconds in which you read the output. No other tab, no phone, just reading.
• Specific check questions. For every email: do the facts match? do the names match? does the tone match? is there something in there I didn't say?
• “What would you write differently?” Once a week, ask a colleague to read one random AI email from that week. That's your second pair of eyes.
• Mail on behalf of a manager double-checked. The impact is biggest there, tone and facts in someone else's name. Write down for yourself the rule: no AI-generated mail going out on behalf of the department head or training supervisor leaves without me, or the manager themselves, having explicitly read it.

Lesson 5.6: What do you do when something really goes wrong?

Goal: a calm step plan for the rare time an AI mistake reaches the outside.

It happens: an email with an invented patient detail, minutes with a meeting decision that wasn't taken, a referral that was a Claude hallucination. The course is meant to prevent this, but nobody prevents 100%. What do you do if it still happens?

1. Breathe in. Don't hide. Sweeping it under the rug is tempting and always leads to bigger problems later. AI mistakes are collective learning material for the department.
2. Determine the scope. Who has seen the mistake? What is the factual error? What needs to be corrected in a follow-up?
3. Short rectification. Email a quick short message: “In my earlier mail/minutes [date] it said [X]; this is not correct. Correct is [Y]. Apologies for the confusion.” Not too long, not rambling. Acknowledge, correct, move on.
4. Discuss it once. Tell your manager and fellow secretariat what happened and how you'll prevent it next time. Not for punishment, for learning.
5. Update your prompt. Which rule would have prevented the mistake? “Don't invent patient names”? “When in doubt: write ‘not specified’”? Add to your library.

Lesson 5.7: Colleagues and team onboarding

Goal: be able to transfer the course you took to a colleague in five minutes.

An often underestimated opportunity: you've been through this course, and there are probably two, three or five colleagues at your secretariat doing the same work. A short team onboarding (~15-30 minutes) can be enough to give everyone the same basis, with the bonus that you all work the same way.

Proposed content for a short team handover:

1. (2 min) The three tools: ChatGPT, Claude, Gemini, what they do for our work.
2. (3 min) Traffic light: what you do and don't paste in a public chat. Short exercise with three examples from your own work.
3. (5 min) Four building blocks: role + context + task + format. First live demo with one real email.
4. (5 min) Template prompt: together you can grab your prompt library.
5. (5 min) What we do when it goes wrong: rectification, learning, moving on. Everyone may know this.
6. (5 min) Questions.

Lesson 5.8: Short check questions for IT/privacy officer

Goal: in one short email get the right information from IT.

The course has given you a feel for what's possible. The definitive boundaries are set by your hospital. Below is a short, neutral email you can send to IT or the privacy officer to clarify the basic route.

Take-home of module 5 (and the course)

Done, fill in the short evaluation

Two minutes, anonymous. Afterwards you'll get access to the assignment book: 15 extra assignments at level to keep practising at your pace with ChatGPT, Claude and Gemini.

Go to the evaluation