Lesson 1.0: What actually is a chatbot (without jargon)?

Goal: in two minutes, understand what ChatGPT, Claude and Gemini run on, without any technical background.

Imagine: a very smart intern who has read all the text on the internet, in libraries and books. He didn't memorise the text the way you would, but he learned patterns in language: which word usually follows which word, which tone fits which context, which structure a clean email has. Ask him something and he predicts, word by word, what a fitting answer would be.

That's all a chatbot does: predicting a likely continuation, word by word, based on what it has seen in training. That explains two things that matter for your work:

• He doesn't know anything “for sure”. He predicts language. Sometimes he predicts something that sounds-like-true but isn't. That's called a hallucination. We'll come back to that in a bit.
• He's not a search engine and not an archive. He has no access to your EHR, no access to your mailbox (unless you give him that), and no knowledge of your specific patient. He only has what you give him at this moment in the chat.

That's it. No artificial intelligence that “thinks”. No magic. But an astonishingly good language assistant, if you give him the right context.

Lesson 1.1: Three tools, three characters

Goal: in one go, feel how ChatGPT, Claude and Gemini differ.

ChatGPT (by OpenAI), Claude (by Anthropic) and Gemini (by Google) are all three chatbots running on so-called large language models. Under the hood they do something similar, but in use they feel clearly different:

This is not a ranking; it's a first taste profile. In the next lesson you'll test it yourself.

Lesson 1.2: Live: same question, three tools

Goal: feel the style differences by asking exactly the same question in three chats.

We're not going to explain this. We're going to do it. Below, open ChatGPT and Claude in separate tabs, and if your work uses Google Workspace: also Gemini. Ask all three exactly the same question. Save the answers somewhere (briefly copy into Notepad or Word).

Lesson 1.3: What you paste, what you don't

Goal: in 30 seconds know whether something belongs in a public chat.

The rule of thumb is simple: public chats are not a hospital environment. For you that means: anything you wouldn't put on a poster in the waiting room without embarrassment, you don't simply paste into ChatGPT, Claude or Gemini. Within that, the traffic light helps:

• Green, public information, fictional names, your own text, an idealised practice case.
• Amber, internal department names without patient data, rosters without identifying combinations.
• Red, patient information, BSN/SSN, EHR exports, a complete letter, unique combinations of diagnosis+age+region.

For amber: anonymise or use dummy names. For red: don't paste, or use a hospital-approved environment (think Microsoft Copilot in your work environment, Claude Enterprise or an EU route via Mistral). In module 5 you'll learn how to apply this systematically.

Lesson 1.4: The four building blocks of a good prompt

Goal: you no longer have to puzzle over “how do I ask this exactly?”

A good prompt almost always has four things:

1. Role, “You are an experienced staff secretary cardiology.”
2. Context, “Our department has outpatient clinic, MDT and training meetings. A colleague is on leave.”
3. Task, “Write an email to the clinic coordinator to hand over this week's agenda.”
4. Format & tone, “Short (max 8 lines), business-friendly, in bullet form.”

Optionally a fifth: constraint: “No patient data; do not name patient-specific appointments.” This is surprisingly powerful to prevent runaway AI output.

Lesson 1.5: Hallucinations: what they are and how to spot them

Goal: know why a chatbot sometimes confidently says nonsense, and how to filter it out before you hit send.

A hallucination is when a chatbot writes down something that sounds correct but is actually invented. An invented phone number, a non-existent NICE guideline, a date that doesn't fit, or a colleague's name that you never mentioned but suddenly appears in the minutes. This happens regularly, not an incident but predictable behaviour of the model.

Why? A chatbot predicts language that fits. If you ask “what is the phone number of the cardiology clinic” and it didn't see that during training, it will more likely invent a plausible-sounding number than say “I don't know”. Unless you explicitly tell it to.

The good news: you can largely prevent hallucinations by putting two sentences in your prompt:

1. “Use only what I tell you here.” (Prevents it from adding its own ideas.)
2. “When in doubt: write ‘not specified’.” (Gives it explicit permission to fill in nothing.)

In module 3 you'll see these two rules return in long-text prompts repeatedly. For now: remember them.

Lesson 1.6: The context window in plain words

Goal: know why a tool “can't do it all at once” and how to work around it.

Every chatbot has a kind of working memory: it “sees” only a limited amount of text at a time within one chat. That's called the context window. Paste more than that at once and it silently forgets the beginning of what you gave it, without always saying so clearly.

The size of this working memory differs*:

* These numbers are a snapshot: vendors enlarge their working memory regularly, so every few months the sizes can look different again. The relative picture (Claude/Gemini roomy, free ChatGPT tight) has so far stayed recognisable.

A “token” is roughly a word (or part of a word). An average A4 page of English text is about 500–700 tokens. So: 30 pages is roughly 20,000 tokens; 100 pages is 60–70,000 tokens.

What this means for your work:

• Short email, short table: all tools work fine.
• An hour-long minutes transcript, a mail thread of 15 messages: free ChatGPT gets tight quickly; Claude/Gemini don't.
• An entire 80+ page ESC guideline in one go: Claude is safer; with ChatGPT you'll need to split.

Lesson 1.7: Your first preference list

Goal: have a first workable choice today, not a final verdict.

Based on the three experiments above, you'll now make a short “who do I pick when” list. It doesn't have to be perfect; you'll adjust it later.

Take-home of module 1