Minutes, agendas and long documents
Here you enter the domain where Claude really shines: long texts. A raw Teams transcript, a mail thread of fifteen messages, a ten-page guideline. You'll learn how to extract usable structure without having to sit through every detail yourself.
Time: ~75 minutes. Prefer working on a laptop, overview helps with long text.
Lesson 3.1: Why Claude for long text (and when ChatGPT is fine too)
Goal: know which tool to open as soon as a document goes beyond two A4 pages.
The most important practical difference between the tools for your work: context window. That's how much text the chatbot can “remember” at once during your conversation. Summarising very long documents goes smoothest with whoever has the largest context window.
| Tool | Practical context window | What fits (roughly) |
|---|---|---|
| Claude Fable 5.1 | 1M tokens (lighter tiers often 200K) | ~250 pages of text, entire guidelines, long transcripts |
| ChatGPT (GPT-5.6) | ~1M tokens | ~800 pages, similar to Claude |
| Gemini 3.1 Pro | 1M tokens (in workflows; 200K-1M in chat depending on product) | Similar to Claude; in Workspace combinable directly with your documents |
For the secretariat this means: if you want to summarise a long Teams transcript, a mail thread of 15 messages or an entire guideline, Claude is often the calmest route. For shorter pieces all three do fine.
Question for lesson 3.1
1. You receive a Teams transcript of an 80-minute meeting. Which tool do you reach for first?
Lesson 3.2: Live: minutes from a raw transcript
Goal: from messy transcript to publishable minutes in one iteration.
A lot of team meetings now run via Teams or Zoom, with automatic transcripts that are messy (wrong speakers, words jumbled, “uh” and “mhm” everywhere). The biggest time win is here, provided you don't forget that a transcript often contains identifying material (speaker names, possibly patient details). Therefore treat a transcript as amber or red until you have cleaned it up.
Practise safely
For this exercise, use the fictional transcript below. Process your real transcripts later in a hospital-approved environment, or after anonymising in a public chat, according to local policy.
Live exercise 3.2: Make minutes from a fictional staff meeting
Open Claude. Paste the entire prompt (including the fictional transcript). Get clean minutes back.
Iteration: if the first minutes sound too formal, ask “Make it shorter and more direct, replace full sentences with single lines where possible.” If it misses action owners, ask “Explicitly add the action owner after each decision.”
Lesson 3.3: Meeting agenda from a long mail thread
Goal: turn a mail tangle into a one-A4 agenda with decision points.
Very classic: a week of back-and-forth emails about the preparation of a meeting, and you need to turn it into an agenda. This is exactly the kind of work Claude (or ChatGPT) takes off your hands fast, provided you structure what you get back.
Live exercise 3.3: Paste your own (anonymised) mail thread
Open Claude. Anonymise one of your mail threads (replace real names with “Colleague A”, “Cardiologist 1”, etc.; replace specific patient items with “Patient case X” or leave them out). Paste into the template below.
No mail thread at hand? First ask Claude: “Invent a fictional mail thread of 8 messages between three cardiologists and the staff secretary about preparing a department meeting.” Paste that output into a new chat and apply the prompt above to it.
Question for lesson 3.3
2. What's the value of an agenda with ‘to decide’, ‘FYI’ and ‘open questions’ separated?
Lesson 3.4: A long guideline on one A4 for your doctor
Goal: turn an 80-page ESC guideline into a one-A4 briefing.
Cardiologists like to read, but one A4 with the main lines before diving into the full guideline is gold. This is something you as a secretary can prepare with Claude (or in Workspace with Gemini). It's a nice service step.
To be fair: as a staff secretary you may not need to do this very often. But the skill you're practising here, loading a long document (such as a PDF) and having it summarised against specific requirements, is one of the most useful AI skills there is. Exactly the same approach works for a policy document, a set of regulations, a long contract or a thick meeting dossier.
Live exercise 3.4: Your own summary prompt for guidelines
Open Claude. Grab a public ESC guideline or a Dutch guideline (NHG, NVVC, NIV). Paste the first 20-50 pages of text into a Claude chat together with this prompt.
Important: you're not producing a clinical summary for decision-making; you're producing a reading prioritisation for your doctor. The doctor decides what to take on board. Module 5 goes further into this boundary.
Lesson 3.5: Claude Projects: your own knowledge base
Goal: know that you can create a “Project” in Claude with its own knowledge base, and when that's worthwhile.
Claude has a feature called Projects. You can “attach” documents to them (PDFs, Word, txt) and after that Claude automatically refers to those documents in every chat inside that Project. For a staff secretariat that's a gamechanger:
- Project “Our department”: hang internal agreements in there (house style, mail tone, list of cardiologists + cover, fixed agenda items). Every question inside this project already “knows” who does what.
- Project “External referrers”: an anonymised list of regular referrers, their preferences, and which cases they often submit.
- Project “Policy and guidelines”: collected notes from IT, privacy and management. Questions about “is this allowed by policy” get a more fitting answer faster.
Important rule
A Project is not a secret space. What you upload is processed by Claude. For consumer Claude: no patient data, no confidential personnel files, no contracts. For Claude Enterprise via your hospital's licence: check with IT what is allowed. For you now: use Projects for general, non-confidential work information.
Live exercise 3.5: Create your first Claude Project
Open Claude. Click “Projects” (or “New Project”) in the top left. Create a Project named “My staff secretariat”. Add this text to the Project's instructions:
Test: start a new chat inside this Project and type one sentence: “Write a short internal email to collect availability for the MDT in two weeks.” Feel how much context Claude already has, you only need to say one more sentence.
Question for lesson 3.5
What's the main advantage of Claude Projects for the secretariat?
Lesson 3.6: Research for the department: a mini overview from multiple sources
Goal: turn four or five sources into a clean one-A4 your doctor or department head can read in five minutes.
Cardiologists and department heads regularly ask secretariats to “find out how things stand” on a topic. Not for clinical decision-making, they do that themselves, but as orientation for a meeting, a training moment or a practical question. This is a nice service step you can do easily with Claude (long context).
Live exercise 3.6: Produce a mini overview
Gather four or five short public texts on one theme close to your work. Two nice options: (a) a public hospital or department protocol (e.g. around onboarding, expense claims or meeting structure) plus a few related information pages, or (b) the PhD regulations of the University of Amsterdam (UvA) plus a few university pages about doing a PhD, handy when a resident or cardiologist starts a PhD track and you want the practical steps lined up. Paste them into Claude. Add this prompt:
Boundary: you don't produce a clinical conclusion, and the cardiologist uses the overview to set their own reading plan, not to decide. Add to the overview explicitly: “Mini overview for reading prioritisation, not for clinical decision.”
Question for lesson 3.6
What's the right boundary for the secretariat on such a mini overview for the cardiologist or department head?
Lesson 3.7: Spotting hallucinations in long text
Goal: pull an invented detail out before it lands in minutes.
With long text there's one risk that's bigger than with short email: AI subtly invents. An invented action owner, an invented page number, a fictional name that only appears in the summary but not in the original. That's hallucination.
The trick to prevent it:
- Explicitly put in your prompt: “Don't invent anything. When in doubt write ‘not specified’ or ‘unclear’.”
- Read the AI output with one question: “Does this also come literally from the source?” If not: ask the AI to add the source sentence.
- For action owners and deadlines: have the AI add ‘[source: timestamp ...]’ in square brackets. When in doubt it writes ‘[source: not specified]’.
Question for lesson 3.7
3. What's the best defence against hallucinations in your minutes?
Take-home of module 3
Long → Claude
At >5 pages or a thick transcript, Claude is often the calmest route. ChatGPT/Gemini can also, but Claude has the most spacious context window.
Minutes format
Date / attendees / decisions + action owner + deadline / open questions. Makes your minutes immediately usable.
Briefing for your doctor
A one-A4 reading prioritisation from a long guideline is a golden service step. But: no clinical advice.
Hallucination guard
“Don't invent. When in doubt ‘not specified’.” Standard rule in all your long-text prompts.
Projects save repetition
Claude Projects holds context, tone and rules. You don't have to re-explain who you are and how you write each time.
Research = reading prioritisation
A mini overview of 4-5 sources for your doctor is a golden service step. Disclaimer: reading prioritisation, no clinical advice.