Use Claude with confidence — in the practice and at home.
No tech knowledge required. No generic AI phrases. This guide shows you visually, in the right order and with over 60 ready-to-use prompts, how to put Claude to work on real everyday medical problems from day one — from patient communication to vacation planning.
"I have never used ChatGPT or Claude. But my daughter says it changes everything. I want to understand where it helps in the practice and in my private life — without a degree in prompt engineering."
Dr. Butze accompanies you through every section of this guide. Every example comes from his real practice or family life.
Three clear stages — no overwhelm
This is how you build confidence step by step, without dealing with every feature at once on day one.
Day 1 — First wins
Create an account, understand the chat window, complete three simple tasks (patient education text, email, research). Pure experimentation.
Go to the getting-started chapterWeek 1 — Projects and routines
Set up your own "rooms" for practice and private life (Projects). Store templates and your own writing style once — Claude remembers.
Discover ProjectsMonth 1 — Your own assistants
Write system prompts, shape your own "staff members" — e.g. a patient-letter assistant, a travel planner, a literature scout.
Build system promptsBe impressed in 5 minutes
Three prompts that show the difference immediately. Just copy, paste into Claude, read the answer.
1. Patient education text in plain language
2. A difficult reply to a Google Maps review
3. Plan a family trip — in one go
You don't need to know how to prompt "correctly." You only need to say what you want, for whom, in what tone, at what length — and which details you typically overlook. Claude takes care of the rest.
The six areas of Claude
That's all you need to understand. Each area has a clear purpose — and its own page in this guide.
Chat
The main window. One question, one answer. Like a conversation with a smart colleague.
Projects
A dedicated "room" for a recurring topic (e.g. practice marketing). Claude remembers guidelines and files.
System Prompts
A kind of "job description" for Claude. Permanently defines style, role, rules, tone.
Artifacts
When Claude builds a "real" document, a form, a table, or a small application instead of just answering.
Practice use cases
Ready-made scenarios for the family practice: communication, research, administration, continuing education.
Private use cases
Family, travel, finances, hobbies, learning — make the invisible tasks of everyday life more efficient.
Never enter real, identifiable patient data into Claude. Not the full name, not the date of birth, not the insurance number. This is non-negotiable — and it isn't necessary either. In every example, we show you how to describe cases so that Claude answers as helpfully as possible without you taking on compliance risks. → Complete rules & examples
Ready for the first step?
Start with the chapter "What is Claude anyway?" — five minutes of reading, and you have the foundation.
What is Claude — and what is it not?
A comparison that actually fits: Claude is like a highly competent, fast, very well-read assistant — who, however, doesn't "practice" any profession and remembers nothing you told him yesterday (unless you tell him again).
"'It's a language model' isn't enough for me. I want to know: What can I realistically expect, where are the real limits, and which terms do I need to know?"
The fitting analogy
Imagine Claude as a young, very smart assistant in a library. Over the past years, this person has read millions of texts — studies, books, journal articles, forums, manuals. He can summarize, draft, structure, compare, and translate at lightning speed.
But: this person has never been a patient himself, has never run a practice, is not licensed to practice medicine — and his library reflects a certain point in time. He never leaves the library unless you send him out (that's "web search"). And at the end of the day he forgets the conversation when you leave the room — unless you have set up a permanent desk for him (those are "Projects" and "Memory").
Claude is ideal for everything you read, check, and cross-verify with clinical experience yourself. He is not ideal as the sole medical decision-making authority. You remain the doctor — Claude is the fast, friendly staff member.
What Claude can do — and what it can't
Claude is very good at…
- Drafting texts (patient letter, email, blog article, patient education text)
- Simplifying complex content — to any desired language level
- Structuring (lists, comparison tables, SOAP notes, pros/cons)
- Translating between languages — including medical texts
- Summarizing long documents (guidelines, PDFs, studies)
- Brainstorming and collecting differential diagnosis ideas
- Building code, tables, small web apps, and visualizations
- Describing and interpreting images (e.g. ECG sketch, photo of a report)
Claude is poor or dangerous at…
- Independent clinical decisions (never without your review)
- "Current" information without active web search
- Guarantees — hallucinations happen, always cross-check
- Real patient data — never enter anything identifiable
- Legally binding statements without medical / legal review
- Image generation (Claude does not generate images, videos, or audio)
- Making phone calls, booking appointments itself (except with connectors)
- Remembering old conversations if you don't enable that
How does it work at a high level?
You enter a prompt — that's simply your question or task. Claude processes this text and responds. You can follow up, correct, go deeper. That's called a conversation.
You type or speak
A question, a task, a piece of text — in plain German or English. Long inputs are fine. Attachments (PDF, images, Word) too.
Claude reads and processes everything in context
The "context" is Claude's short-term memory within the current chat. Very large, but limited. As long as the conversation is running, Claude can access everything that came before.
Claude responds — and you iterate
The first answer is rarely perfect. Just say what you want to change: "Shorter." "At B1 level." "With three examples." "In a different tone." That's the lever.
Optional: add tools
Web search, file upload, connection to Google Drive, calendar, Notion, and more (so-called "connectors"). With these, Claude can also do things instead of just answering.
Mini glossary — the eight words that keep coming up
Prompt
Model (Sonnet, Opus, Haiku)
Opus is the strongest and slowest,
Sonnet a good all-rounder, Haiku the fastest for light tasks.
For doctors, the default setting is enough — you don't have to actively choose.
Context window
Hallucination
System Prompt
Project
Artifact
Connector / MCP
Which model should I choose?
For 95% of all tasks in the family practice and in private life, Claude's default setting is the right one. You don't need to change anything. If you want the details:
Sonnet
The fast all-rounder. Ideal for emails, patient education texts, research, brainstorming, daily planning. Default in Claude Pro.
Opus
The strongest model. Worth it for long, complex tasks — e.g. when you want a 60-page guideline summarized or a complex letter drafted.
Haiku
The fastest, lightest model. For very short, quick requests — such as "What's that medication called again with active ingredient X?".
Just leave it on the default setting. Only switch if you notice that a task really needs long answers or particularly difficult reasoning — then pick Opus.
The chat — your constant sparring partner.
The chat window is the fastest way to get something done with Claude. You open a new tab, write your question, get an answer, refine. Here you'll learn what separates a good prompt from a bad one — and collect ten templates that fit straight into your daily routine.
When is the chat the right tool?
Chat is right for…
- One-off, quick requests
- Quick research and "what if" questions
- Spontaneous patient communication
- Ad-hoc translations or explanations
- Summaries of long texts
- Brainstorming and differential diagnosis ideas
For recurring tasks, better to…
- Create a Project (see chapter)
- Define your own system prompt
- Upload important files permanently
- That way Claude doesn't need to be re-briefed every time
The anatomy of a good prompt
A good prompt implicitly answers five questions: who (role), what (task), for whom (target audience), how (tone & format), what to watch out for (constraints). You don't have to phrase it in this order — but the more of this is in the prompt, the better the answer.
Bad → better → very good: the same task in three stages
Role · Task · Target audience · Tone · Format · Constraints. If you keep these six building blocks in mind, you'll automatically write better prompts — without tricks and without a prompt engineering course.
Ten ready-made prompts — straight from Dr. Butze's daily routine
You can copy each of these prompts as-is, adapt it to your situation, and paste it into Claude.
1. Explain a test result in plain terms
"Ms. M. took her lab report home but doesn't dare to ask. I want to explain to her in one sentence that her thyroid values are borderline."
2. Get a quick overview of a current guideline
"I don't have the DEGAM guideline on irritable bowel syndrome at hand and need the key points for a consultation in 20 minutes."
3. An empathetic refusal to a pharma rep
"I received three appointment requests from pharmaceutical reps today. I want to decline all three politely but firmly."
4. Differential diagnosis brainstorming
"I have an unclear case — a 38-year-old patient with recurrent headaches, moderate nausea, mild visual disturbances. I want to be sure I'm not missing an avenue."
5. Birthday speech for the father-in-law
"My father-in-law is turning 75. I have to give a short speech — and can't find a way to start."
6. Prepare for an insurance comparison
"I have three private health insurance offers on my desk. I want to understand them before I talk to the broker."
7. Revive a language
"I learned Spanish at university. I want to reactivate it on the side without an app — 10 minutes a day."
8. Draft a job posting for a medical assistant
"I need a new medical assistant and can't find one. Maybe it's also because of my old ad."
9. Think through a parenting question
"My son (11) wants a smartphone. I'm unsure. I need clarity before we talk about it."
10. Summarize a guideline PDF
"I downloaded a 60-page DGIM guideline. I need it as an update in 5 minutes."
Common beginner mistakes and how to fix them
"It produced nonsense." → You gave too little context
"It was too long." → Set a word limit
"It sounds like AI." → Say what it should not sound like
"The facts were wrong." → For medical facts, enable web search or prompt differently
"It assumed too much." → Stop it and ask
Iteration: 70% in one minute → 95% in ten
Don't expect the first answer to be perfect. It rarely is. But: with 4–5 short follow-up comments you get extremely close to the goal. Example iteration:
It's not "the perfect prompt." It's the willingness to iterate in 3–4 steps. Once you internalize that, you have truly understood Claude.
Projects — your own rooms for recurring topics.
A chat is like a single conversation in an elevator. A Project is like your own consultation room: the same style, the same documents on the table, the same rules — every time you enter it. You don't have to explain everything each time.
"Every time, I write to Claude all over again: 'Answer in my practice's style, briefly, on friendly terms with patients, always with a note about follow-up consultation, no jargon.' That gets annoying after the tenth time. There must be a way to store this once."
There is. It's called a Project.
Chat vs. Project — the decisive difference.
Project
- For recurring topics ("patient communication")
- Its own instructions, stored once
- Its own files (e.g. practice guidelines), Claude always has them "on the table"
- All chats within it share the same guidelines
- You open the Project — and start working immediately
Regular chat
- One-off, isolated questions
- You have to provide context every time
- No "personal library"
- Every chat starts from zero
- Good for spontaneous things, bad for recurring ones
Important: within a Project you still have any number of individual chats. The Project is not "one single, everlasting conversation," but a frame in which all conversations share the same rules and the same background.
Three building blocks — that's all it takes.
Project instructions ("ground rules")
A text field in which you permanently tell Claude how to behave in this Project: role, tone, language, length, formats, taboos. This is not a one-off prompt — Claude automatically reads these instructions in every chat in the Project. A good project instruction text is 100–400 words long.
Knowledge base ("library")
Files that Claude has "on the table" in every chat of the Project: your practice guidelines, your own style guide, FAQs, templates, patient education texts, a vocabulary list. With Pro/Max, large files are automatically made searchable (RAG).
Chats ("the actual work")
Every conversation you have in the Project is visibly saved. You can jump back at any time. However, different chats don't share memories — each chat starts "fresh" with the background knowledge from building blocks 1 and 2.
Project instructions are a special form of "system prompt," just project-specific. They apply in addition to your profile settings (which apply to everything). If both exist, they complement each other. The profile settings say "this is who I am," the project instructions say "in this room I work on X."
Create a Project — in 4 clicks.
Left sidebar → "Projects"
In the left column, click the Projects menu item (or
go directly to claude.ai/projects). You'll see either an empty
overview (normal at the beginning) or your existing projects.
Top right → "+ New Project"
Enter a name ("Patient communication family practice") and optionally a description. Both are only visible to you and have no effect on Claude.
Enter the project instructions
On the Project's main page you'll see an input field "Instructions" or "Set custom instructions". This is where the text goes that we prepared in the templates below — copy & paste.
Fill the knowledge base (optional, but strongly recommended)
On the right side of the Project main page there is an area called "Project knowledge". Use the + icon to upload files (PDF, DOCX, TXT, CSV — up to 30 MB per file). Examples below.
Now click into a chat within this Project. Ask a question that fits the context — and watch how Claude automatically takes your guidelines and files into account without you having to repeat anything.
Five ready-made Project templates for doctors
Each template consists of a name, a project instruction (to copy), and a suggestion for which files belong in the knowledge base.
🩺 Patient communication family practice
For all patient education texts, patient information, reply emails, and explanatory letters. Claude speaks here the way you would write — if you had unlimited time.
Project Instructions (copy)
Knowledge base — what to upload?
- Your existing patient education texts (as style examples — anonymize first)
- A list of "forbidden terms" you never want to see in patient mail
- Optional: a file with common patient questions and your standard answers
📚 Continuing education & literature overview
Summarize studies, understand guidelines, compress new recommendations into two paragraphs. Here, Claude becomes your personal, critical editor.
Project Instructions (copy)
Knowledge base — what to upload?
- PDFs of current guidelines (DEGAM, AWMF) — where publicly available
- A list of your clinical focus areas ("cardiovascular risk", "type 2 diabetes", "early dementia")
- Your own notes from previous training courses as "vocabulary"
📣 Practice external communication
Website, Google profile, newsletter, answering machine scripts, job postings, notices — you want to come across as a modern, attentive practice, not like 1998.
Project Instructions (copy)
Knowledge base — what to upload?
- Your current website as a PDF or screenshots
- A list of practice services, office hours, team
- Examples of texts from other practices that you like (URL as a note)
- A short "brand DNA" (10–15 keywords: "time, listening, honest, Berlin, family doctor, local, no esoteric nonsense…")
🌍 Family life & travel
Vacation planning, weekend activities, birthday planning, family calendar, school communication. The Project area where you're not a doctor but a father or partner.
Project Instructions (copy)
Knowledge base — what to upload?
- A list of the most important family dates in the year (school vacations, birthdays, holidays)
- A short note with favorite places ("We love South Tyrol, the Baltic Sea, the Black Forest, Italy…")
- The family's allergies / intolerances
- A few example photos with daily plans that worked well in the past
🏠 Authorities, insurance, contracts
Letters to the health insurance fund, tax advisor, insurance company. Understanding contracts. Responding to outrageous demands factually but firmly. The area that drains energy every Sunday evening.
Project Instructions (copy)
Knowledge base — what to upload?
- Your current letterhead / address
- A list of the file and contract numbers you regularly deal with
- A "hall of fame" of your own letters that once worked very well
Five rules of thumb for good Projects
1. Better 5 specific ones than 1 big one
One Project per clearly delimited context. "Practice communication" and "family" don't belong in the same Project, otherwise tone and style blur.
2. Project instructions are a living document
Update them every 3–5 chats. Whenever you notice: "I don't like this" — add it to the instructions. It keeps getting better over time.
3. Prohibitions are often more effective than commands
"Never jargon without translation," "Never longer than 200 words" — such negative rules often work better with Claude than positive guidelines.
4. Keep the knowledge base lean but current
10 clean, well-named files are worth more than 50 chaotic ones. Tidy up every 1–2 months.
5. Never identifiable patient data
Not in the knowledge base either. Create example files with fictional people. Details in the data protection chapter.
6. Pro / Max / Team supports RAG
With Pro+, Claude can find what's currently needed even in very large knowledge bases. On Free, the knowledge base is smaller — still enough for beginners.
System prompts — onboard Claude, once and for all.
A system prompt is the "job description" for Claude. A kind of onboarding letter to a new employee that tells him: "This is how you work for me. This is the role. These are the rules. This is how I speak. You never do this. You always do that." Written once — permanently effective.
"There are two kinds of answers I get from Claude. Answers I can use immediately — and answers I first have to rewrite into 'my language.' I want there to be only answers of the first kind."
That is exactly what system prompts are made for.
Three types of instructions — and why that matters.
Profile settings
Apply everywhere. Here you tell Claude who you are and how you work — regardless of the topic. Example: "Write concisely, avoid clichés, I'm a doctor".
Project Instructions
Apply only within one Project. Here you tell Claude how to behave for a specific topic — e.g. "patient communication". We covered this in the Projects chapter.
Chat prompt
Applies only to this one chat. This is your actual question. It supplements the two levels above without overriding them.
Claude always reads all three — profile first, then Project, then chat. This gives you the best of all worlds: permanent ground rules + context-specific guidelines + spontaneous request.
The profile settings — your universal system prompt
In the settings at the top right → "What should Claude call you?" and "What should Claude know about you?". This is where the most important text you will write goes — it applies to every conversation, every Project, every context. A ready-made, well-functioning template:
Do not put project-specific things in the profile settings, like "You are a marketing copywriter for my practice." Otherwise that also applies when you ask for restaurant suggestions — and Claude will answer in a confused way. Project-specific content goes into project instructions, not into the profile.
Seven building blocks that make up every good system prompt
You don't have to use all seven. But the more of them you cover, the more precise Claude's behavior becomes. Order in the system prompt: in this sequence.
Role & purpose
"You are…" — a clear functional description. "You are my co-author for patient education texts in a family practice." Avoid overly generic roles ("You are a helpful assistant" doesn't help Claude).
Target audience / context
Who is Claude actually writing or thinking for? "The recipients are patients between 30 and 80 years old, mixed education, many read quickly and on the side." Without this point, the wrong tone is the most common result.
Tone & style
Concrete and enforceable. Instead of "friendly," better "polite but direct; active verbs; sentences under 18 words; no advertising clichés". Better 5 specific rules than 20 vague ones.
Content rules (dos)
What Claude should always do content-wise. "Every text ends with a clear next step for the patient." "With medical recommendations, always add 'if in doubt, please contact the practice first.'"
Content taboos (don'ts)
Often more important than dos. "Never specific dosages without my explicit instruction." "Never make diagnoses." "No promises of cure." "Never jargon without translation."
Format standards
Length, structure, language. "Standard length: 120–180 words, unless I explicitly request otherwise. Letters: salutation, body, closing 'Kind regards, your practice team Dr. Butze'."
Self-check before every answer
A small "pre-flight checklist" that Claude mentally runs through before every answer. "Did I make a claim I cannot support from the source?", "Did I use jargon without translation?", "Does the text end with a next step for the patient?" — Three such questions noticeably improve quality.
Claude writes your system prompt — for you.
You don't have to formulate all of this yourself. One of the most elegant applications: Let Claude draft the system prompt for you. The following meta prompt — just paste it into a regular chat — delivers a tailored proposal.
You provide the context, Claude leads the conversation. After 5–10 minutes of dialogue you have a system prompt that fits you — much better than any generic template. If it's not perfect, just say: "Tighten point 3, it's too soft for me. Add a point about X." Iteration is part of the method.
Six ready-made system prompts
Each one is ready to use immediately. Copy, paste into your profile or a Project, adjust slightly in one place — done.
Patient education text assistant
For all patient information, package insert explanations, procedure consent information, and written explanations.
Referral letter drafting helper (fully anonymized)
Important: this prompt works exclusively with fictional or anonymized cases. Real patient data doesn't belong in Claude (except Claude for Work with a data processing agreement). Read the data protection chapter.
Literature & guideline summarizer
For weekend continuing education. You provide a study or guideline as a PDF / text / link, Claude delivers the GP-relevant summary in half a page.
Practice marketing copywriter
Website, Google profile, notices, job postings, answering machine scripts. HWG-aware, without advertising clichés.
Family travel planner
For vacation planning with the family, weekend trips, travel logistics.
Letters to authorities & insurance companies
Factual, precise letters — polite, but firm. Never threatening, never submissive.
System prompts are never finished.
Your first system prompt will be okay. The third will be good. The tenth will surprise you. Three tips for iterating:
1. Note what annoys you
Every time you have to rewrite an answer — briefly note what didn't fit. After 5 such moments you have 5 new rules for your system prompt.
2. Prohibitions are often more effective than commands
"Never use the subjunctive" works better than "Write concretely." Negative anchors hit Claude more precisely.
3. Claude improves its own prompt
Say: "Here is my current system prompt. Read it, then show me five concrete weaknesses — and make one improvement suggestion for each."
System prompts are what make the difference between "Claude is a nice toy robot" and "Claude is my most reliable staff member." Every good workflow stands or falls with a good system prompt.
Artifacts — when Claude builds an attachment.
Sometimes Claude answers in the chat. And sometimes a second window opens to the right with a finished document, a checklist, or a diagram — cleanly formatted, downloadable, further editable. Those are Artifacts. They change how you work with Claude.
"I asked Claude to create a checklist for the U9 well-child checkup. Suddenly a separate document appears to the right of the chat — formatted, with checkboxes, ready to download. That was new. And exactly what I needed."
Think of Artifacts like an attachment in an email.
In the chat window you talk to Claude — like in a chat. But when Claude hands you a document — a checklist, a letter, a table, a diagram — it doesn't land in the conversation but in a separate window to the right. That is the Artifact.
So the document stays readable. A twenty-page letter in a chat line would be illegible. A diagram in plain text would be nonsense. Artifacts are Claude's way of saying: "This is no longer a conversation — this is a result. Look at it, change it, download it."
Chat answer
Short, conversational, gets lost in the conversation. Three paragraphs about the pros and cons of a therapy. You read, reply, move on.
Artifact
Standalone document. You can copy it, download it, forward it — and in the next message say: "Make the section on aftercare longer." Claude changes only that section.
When Claude creates an Artifact — and when not.
Claude decides on its own. But there are clear patterns. If your request has one of the following characteristics, the chance is high that an Artifact is created:
Length
The result will be longer than ~15 lines or 1,500 characters. Example: "Write a two-page patient letter."
Structure
You ask for a clearly structured document — checklist, table, letter, report. Example: "Create a checklist for the hygiene training."
Reuse
You want to reuse the result later, copy it, or save it as a file. Example: "Make a template for the patient consent form."
Visualization
You want a diagram, a table, or a process plan. Example: "Show me the treatment process as a flowchart."
Simply write: "Make this as a document." or "Create this as an Artifact to download." Claude understands that.
Five types of Artifacts that are relevant for you.
Letters, reports, patient education texts
This is the most common type. Cleanly formatted with headings, paragraphs, lists. You can download it as Word or Markdown, copy it, reuse it.
Structured lists, comparisons
When you compare or list something — drug interactions, therapy options, diagnostic differentiation — a table often results. Readable, copyable into Excel.
Flowcharts, process plans
"Show me the U9 checkup process as a diagram" — Claude builds a visual schema that you can print or embed in a presentation.
Calculators, quizzes, interactive aids
Claude can also build simple interactive tools — for example a BMI calculator for the practice website or a small symptom self-assessment tool. Usable directly in the browser.
Simple HTML pages, templates
For your own practice website or a letter with letterhead — Claude can output a finished HTML page. You see the page rendered on the right, can test it and download it. That's the pro level; text Artifacts are entirely sufficient to start with.
Artifact prompts to take with you right away.
Copy, paste, adapt. All prompts reliably trigger an Artifact because they are structured enough.
1 — Patient info sheet for diagnosis XY
A one-page info sheet you can give to patients with a new diagnosis. Understandable, reassuring, without jargon.
Claude can write something wrong or imprecise even with the classics. You remain responsible for the medical content.
2 — SOP checklist for the practice
A reusable checklist — for example for the monthly hygiene training, the emergency equipment, or restocking the examination rooms.
3 — Treatment pathway as a flowchart
A visual process — good for patient communication, staff training, or as a notice in the waiting room.
4 — Packing list for the family vacation
A complete packing list sorted by person — created once as an Artifact, adjusted once, and it's set for the next ten vacations.
5 — Weekly menu planner as a table
Seven days, three meals, with a shopping list at the end — exactly what's annoying on Sunday evenings.
6 — Small BMI calculator for the practice website
A simple, interactive tool that you can embed on your practice website. Patients enter height and weight and see the result immediately.
Claude writes the code. You see the finished tool directly in the Artifact window, can try it out and download it. Your webmaster or practice service provider gets the file and embeds it.
Working with Artifacts — editing, sharing, saving.
-
Improve iteratively
Just say: "Make the 'Aftercare' section longer." or "Remove the medical vocabulary from point 3." Claude changes only what's requested, the rest stays in place.
-
Download or copy
Via the menu at the top right of the Artifact: download as Markdown, Word, PDF. Or put the entire text into the clipboard with one click and paste it into your practice management system.
-
Share with a link
You can also send Artifacts to colleagues via a link. Careful: the Artifact is then publicly visible to anyone with the link — never share patient-related content.
-
Reference it in the chat
You can keep referring to the Artifact during the ongoing conversation: "Write a shorter version of the patient letter for our email reply." Claude knows which document is meant.
Three pitfalls you should know about.
This is how it works well
- Say clearly what should be in the document — structure, length, target audience.
- Have it changed iteratively ("rewrite only section 2").
- Read everything through before sending.
- Anonymize patient data before entering it.
This is how it goes wrong
- Forwarding the Artifact unreviewed to the patient inbox.
- Trusting "it's probably right" for medical content.
- Sharing Artifact links carelessly (public!).
- Mentioning real patient names in the prompt.
Everything you write in the prompt — even if the result ends up in the Artifact — is sent to Anthropic's servers. Patient data doesn't belong there. For details, see Data protection & limits.
From here on it gets concrete: on the next pages you'll see ten professional and ten private use cases — each with a ready-made prompt.
Ten use cases for the practice.
From the patient education text to the letter to the authorities — ten tasks that come up in every family practice, each with a copyable prompt. Skim through, pick out what you need today.
1. No real patient names, addresses, dates of birth, insurance IDs, or findings in the prompt. 2. Review professionally before every send — Claude can be professionally wrong. 3. No medical data on the Free/Pro/Max plan. Details: Data protection & limits.
"I used to spend two more hours writing letters after closing. Now I write a prompt, read the answer, change two sentences. Ten minutes. The rest of the evening I see my kids."
- Rewrite a patient education text in patient-friendly language
- Answer a patient email politely and clearly
- Politely decline a pharma rep request
- Summarize a guideline on one page
- Condense a study/article for daily practice
- Copy for the practice website
- Write a job posting for a medical assistant
- Answering machine scripts for the practice
- Structure continuing education notes
- Letter to authorities (KV, building office, health office)
1 — Rewrite a patient education text in patient-friendly language
The official consent forms from the professional associations are legally bulletproof — but incomprehensible to many patients. Claude turns a technical text into a patient version that you attach as a supplement.
"I can't spend twenty minutes explaining to every patient what a colonoscopy is. Now everyone receives, in addition to the formal consent form, a one-page comprehension sheet — and the conversations become shorter and better."
The patient version does not replace the legally required original consent form. It is a comprehension aid — and that should also be clearly noted on the page.
2 — Answer a patient email politely and clearly
Patients often write long and emotional messages. You reply briefly and factually. Claude bridges this gap without sounding cold.
"I decide what the content is. Claude makes sure it doesn't sound like an administrative notice."
3 — Politely decline a pharma rep request
Four requests per week. Answering each one individually costs time. With a Project and a system prompt, a matter of thirty seconds.
If this comes up more often — which is usually the case — create a Project "External communication" with a system prompt. Then in the chat, this is enough: "Refusal pharma rep." Claude knows the rest. See Projects.
4 — Summarize a guideline on one page
An S3 guideline often has 200 pages. You have ten minutes. Claude turns the complete guideline into a practice summary — which you cross-check against your memory of the original guideline.
Helpful, but no replacement for the original. For every concrete patient case that depends on it: open the guideline yourself, look up the relevant passage. Claude can misinterpret or omit details.
5 — Condense a study/article for daily practice
You have a study from the Deutsches Ärzteblatt or the NEJM open. Forty pages. Question: Does this change how I practice on Monday?
6 — Copy for the practice website
Most practice websites read as if they were built in 2008 by the medical assistant's son. With an hour of Claude work, you get a complete body of copy for a professional practice site.
7 — Write a job posting for a medical assistant
Job postings that work don't sound like administration — they sound like a place where people want to work. Claude can do that if you honestly describe what your practice stands for.
8 — Answering machine scripts for the practice
One script for normal office hours. One for vacation time. One for Wednesday noon. One for emergencies. All written cleanly once, easy to speak, without stumbling sentences.
9 — Structure continuing education notes
Three days of conference, twenty-five talks, a hundred handwritten notes. Claude turns them into something you'll still understand in four weeks.
You can simply photograph your handwritten notes and upload the photo to Claude. Claude reads the handwriting, converts it into text, and structures it according to the scheme above. Works surprisingly well — as long as the handwriting isn't completely illegible.
10 — Letter to authorities (KV, building office, health office)
A letter that sounds appropriately formal, is factually precise, and promises nothing you can't keep.
Bookmark this page. The first time, each case may take ten minutes — by the third time it's three. The time you gain is worth every learning curve.
Ten use cases for private life.
Travel planning, letters to authorities, homework help, birthday poems, weekly menus. Ten everyday situations in which Claude gives you back two evenings per week — each with a ready-made prompt.
"My wife asked me what I was doing on the laptop. I said: 'I'm planning the vacation.' She replied: 'You never plan the vacation.' Forty minutes later we had a complete travel plan for South Tyrol. Since then I've also planned the toy-themed birthday party. And the weekly menu."
- Plan a complete family vacation
- Weekly menu plan with shopping list
- Homework help for the kids
- Gift ideas for difficult occasions
- Write a birthday speech or poem
- Letter to the tax advisor or insurance company
- Write a car sale listing
- DIY instructions step by step
- Letter to the school / class teacher
- Hobby research (wine, bicycle, travel)
Here you can think more freely. But: even privately, don't enter bank PINs, tax IDs, passwords, or your family's health data into Claude. That remains a hard line — see Data protection & limits.
1 — Plan a complete family vacation
From destination to daily schedule. Claude delivers a complete draft, which you then adapt to your preferences.
"I hate travel planning. My wife hates that I hate travel planning. Claude solved our marital problem."
Claude knows many places from its training data — but restaurants close, opening hours change. Before booking, briefly verify online or ask Claude to use web search (feature can be enabled in Claude.ai).
2 — Weekly menu plan with shopping list
Sunday evening, 7 p.m. The same old question: "What are we cooking next week?" Claude answers it in thirty seconds.
3 — Homework help for the kids
You're a family doctor, not a math teacher. But your eight-year-old daughter needs help with fractions. Claude explains it the way a good teacher would — patiently, child-appropriate, without doing the work for them.
If you sit next to your child and they type in the answers themselves, they learn not only the material but also good habits with AI: asking follow-ups, not believing everything, trying their own solution first.
4 — Gift ideas for difficult occasions
Mother-in-law turning 75. A colleague retiring. The eleven-year-old son of an acquaintance has a birthday. Claude doesn't deliver run-of-the-mill ideas, but concrete suggestions if you describe things concretely.
5 — Write a birthday speech or poem
Tomorrow you have to give the speech at your parents' golden wedding anniversary. Or a poem for your daughter's birthday. Claude doesn't write for you — Claude delivers a good draft that you bring into your voice.
6 — Letter to the tax advisor or insurance company
You want to be clearly understood — without hours of back and forth by email.
No tax ID, no account balance, no insurance number. Keep it general — you fill in the letter with the concrete numbers yourself before you send it.
7 — Write a car sale listing
A listing that doesn't look like an "amateur private sale" — but factual, trustworthy, complete.
8 — DIY instructions step by step
Hanging a lamp. Connecting the washing machine. Drilling into a concrete wall. Claude gives you careful, correct instructions — and says clearly where you should call a professional.
Claude knows the theory. But your old building, your specific wiring, your actual fuse situation — only you and an electrician know those. When in doubt: read the instructions, then call.
9 — Letter to the school or class teacher
Sick note covering several days. Request for an excused absence for a family occasion. A constructive reply to a school notice you don't agree with.
10 — Hobby research (wine, bicycle, car, travel)
You've been semi-seriously interested in wine for two years. But you don't want to become a sommelier — you want to know which Pinot Noir to bring to the next family dinner without embarrassing yourself.
If you honestly describe where you stand — beginner, intermediate, advanced — you don't get a textbook answer but a personal recommendation. Works for wine, bicycles, cameras, books, travel destinations, whisky, classical music, hiking routes.
Data protection, GDPR, and the limits of Claude.
This page is not optional. If you work with Claude in the practice, you are responsible — for the data, for the correctness, for the consequences. Here is what you need to know. Clear, without legal waffle, but precise.
Never enter patient data into the consumer versions of Claude. Free, Pro, and Max are not intended for anything patient-related — not even "just briefly", not even "just this one case". Period.
"I had the idea of having Claude type up a referral letter — after all, I had only entered the diagnosis and a few keywords, no name. My data protection officer called me before I could hit 'Send'. Patient data is not just name and date of birth. That was an important lesson."
- Which Claude plan is suitable for what?
- What happens to my inputs?
- What you must never enter
- Anonymizing — the practical guide
- Understanding hallucinations
- Your medical responsibility remains
- EU AI Act — what applies from August 2026
- Pre-use checklist before every use
- When the data protection officer must be involved
Which Claude plan is suitable for what?
Anthropic offers different plans with different contractual assurances. For use with patient data, the difference is existential.
Free, Pro, Max
Personal accounts. No data processing agreement (DPA) available. These plans are aimed at private individuals and general use.
Unsuitable for patient data. Even if you turn off training (which is possible — see below), there is no contract that fulfills the GDPR requirements for a data processor.
Allowed: Everything without patient reference. Practice texts, marketing, letters, research, private matters.
Claude for Work (Team, Enterprise)
Business accounts with DPA option. For these plans, Anthropic offers a data processing agreement (DPA — Data Processing Addendum) and additional security and compliance options.
Even here: The DPA alone does not automatically make use with patient data permissible. It additionally requires your own data protection impact assessment (DPIA), the OK of your data protection officer, and sensibly a restriction to clearly defined use cases.
Recommendation: If you want anything patient-related, Team/Enterprise is the minimum.
Terms, contracts, and options change. Before you set up a business account, have your data protection officer review the current contract terms and the supplementary guarantees from Anthropic.
What happens to your inputs?
-
Transmission to Anthropic's servers
Everything you type into the chat — and all files you upload — is transmitted encrypted to Anthropic's servers. Anthropic is a US company headquartered in San Francisco. Server locations and sub-processors are defined by the current state of the contract documents, which are regularly updated — ask your data protection officer for the currently applicable versions.
-
Processing by the model
The model processes your input and generates a response. During this processing, your data is briefly present in the working memory of the servers.
-
Storage of the conversation (chat history)
Your chats are stored in your account so you can refer back to them later. You can delete chats — that deletes them in your account. How long they remain on the servers in backups or logs is governed by the contract.
-
Training: off by default, but check
Anthropic states that inputs from consumer accounts are not used to train the models without explicit consent. In the settings, a corresponding option is usually switched off by default. Check this in your account: Settings → Privacy → Help improve Claude. Turn it off deliberately — document it in writing.
They are just not baked into future model versions. But they reside on US servers. They are technically accessible — to Anthropic, to sub-processors, in the extreme case to US authorities under the CLOUD Act. That is the basis for the strict rule above.
What you must never enter.
This list is not exhaustive — but if any of it applies, it's clear: not into Claude.
Personal patient data
- Name, date of birth, address
- Insurance ID, KV number
- Phone, email
- Photo, ID copy
Medical findings with personal reference
- Original referral letters
- Lab results, imaging
- History notes with identifying reference
- Also indirect identification (rare + region + diagnosis)
Your own sensitive data
- Passwords, PINs
- Bank details, account statements
- Tax IDs, social security
- Confidential material from ongoing proceedings
Practice-internal confidentiality
- Staff personnel files
- Salaries, contracts with identifying reference
- Confidential correspondence
- Practice management system data
Often not enough. A 47-year-old female patient from Berlin-Charlottenburg with a rare diagnosis is identifiable — even without a name. GDPR-relevant is anything that can identify a "natural person", alone or in combination. The rule of thumb: when in doubt, it counts as personally identifiable.
Anonymizing — how to do it right.
For many tasks you don't need patient data at all. You need the structures. This is how you phrase requests without patients becoming identifiable:
Not like this
- "Write a letter to Ms. Schmidt, born April 12, 1962, regarding her diabetes…"
- "Explain this finding to me: Pat. K.M., 67 years old, atrial fibrillation, Marcumar since 2019…"
- "Here is the referral letter from oncology — can you translate it into plain language for the patient?"
Better like this
- "Write an example letter to a typical diabetes patient switching their medication — placeholders for salutation and date."
- "Explain the finding 'paroxysmal atrial fibrillation on Marcumar' in general terms — when is which adjustment common?"
- "Here is an anonymized example report (markers like [DATE], [PATIENT], [DOCTOR]) — please write a patient-friendly version with the same placeholders."
The three anonymization techniques
Placeholders
Copy the original text, replace personal references with markers:
[NAME], [DOB], [ADDRESS].
The markers remain in the result — you only enter the real data
manually at the end.
Generalization
Instead of "patient K.M., 67 y/o, from Berlin-Charlottenburg, with X", better "a 65-year-old patient with X". Specific identifiers out, clinical constellation in.
Hypothesization
"Imagine the following case: …" — this turns a concrete case into a teaching example. The model responds to the example; your clinical reality stays out of it.
A patient education text is general. A practice letter template is general. A guideline summary is general. Patient data is almost never necessary — it only creeps in because we're used to thinking in concrete cases.
Hallucinations — when Claude is convincingly wrong.
Claude is a language model. It generates plausible text based on statistical patterns in training data. It doesn't know anything in the human sense. This has a specific consequence: sometimes Claude invents facts. Very convincingly.
- Invented study citations ("Müller et al. 2021 showed…") — with author, journal, page number, all freely invented
- Wrong dosage information — plausible, but outside the guideline
- Wrong drug interaction information
- Wrong guideline versions or outdated recommendations
- Invented ICD or OPS codes
- Wrong billing codes (GOÄ/EBM)
How to spot hallucinations
-
Strikingly specific citations
If Claude names a study with author, journal, year, and page — and it sounds important — google it. Only then cite it or act on it.
-
Answers to difficult questions that are too smooth
Complex medical questions often have no unambiguous answer. If Claude nevertheless sounds like "The recommendation is…": become skeptical.
-
Concrete numbers without context
"The success rate is 73%" — where does this number come from? From which study, which patients, which endpoints? Ask.
-
Confidence is not an indicator of correctness
Claude phrases wrong statements just as assuredly as right ones. You cannot infer correctness from the tone of the answer.
— "For medical statements, please always indicate
the uncertainty: 'established' / 'likely' / 'unclear'."
— "If you're unsure about a source, say so openly instead of
guessing."
— "Never invent study citations or guideline references. Better
'I have no concrete source for this'."
Your medical responsibility remains — in full.
Even if you use Claude for preparation, for research, for writing — the legal and professional responsibility for what reaches your patient stays with you. Always.
Malpractice liability
If a patient is harmed by an incorrect treatment recommendation, you are liable. Not Anthropic. Not the model. You cannot claim that the AI said so.
Medical confidentiality (§ 203 German Criminal Code)
If you give patient-related information to an AI system not covered by contract, that can be a violation of medical confidentiality — punishable by law, independent of the GDPR.
GDPR fines
A GDPR violation carries fines of up to 20 million euros or 4% of annual revenue (for SMEs typically much less, but five to six figures are realistic). On top come possible damage claims from individual patients.
Professional consequences
The state medical associations can take action for gross violations of the professional code and data protection — up to and including questions about your license to practice in extreme cases.
You get a lot of things faster and better. But proofreading, checking, signing — that's you. Always.
EU AI Act — what you need to know from August 2026.
The EU AI Regulation (EU AI Act) has been in force since August 2024 and is taking effect in stages. Most general obligations for using organizations apply from August 2, 2026.
Risk classes
The AI Act classifies AI systems into risk classes. AI in the medical field can range from "limited risk" to "high risk" depending on the type of use. Language models that make or prepare diagnostic or treatment decisions tend to be high risk.
Transparency obligations
Patients must be able to recognize when they are interacting with an AI. AI-generated texts that go to patients should be identifiable as such — the legal line here is still in motion.
Human in the loop
For medical decisions, "human oversight" is mandatory. An AI may not make a medical decision autonomously or prepare it alone — the doctor must review, approve, take responsibility.
Documentation
Practices that use AI substantially in patient-related processes will have documentation and evidence obligations. Which ones exactly and from what threshold — is still being specified in detail in guidance from the supervisory authorities.
Much will be specified through implementing regulations and guidance from the EU and national supervisory authorities. Keep an eye on the topic — via your data protection officer, your medical association, the KV.
Ten-second checklist before every Claude use.
Train yourself to mentally run through these five questions before every prompt. After two weeks it becomes automatic.
-
Is there a personal reference in it?
A name, a date of birth, an address, a photo, an insurance ID, a rare diagnosis with region — anything that makes a specific person identifiable? If yes: STOP. Anonymize or don't enter it at all.
-
Am I on the right plan?
Anything patient-related only in Claude for Work with a DPA and after data protection officer approval. General tasks work on any plan — as long as there's no personal reference.
-
Have I disabled training?
Settings → Privacy → "Help improve Claude" → off. Check it once, done.
-
Do I understand what Claude will answer?
If you cannot judge the topic professionally, you cannot judge the answer either. Then Claude is the wrong tool for this case.
-
Do I have a strategy to verify the answer?
Comparison with a guideline? Cross-check with colleagues? Your own expertise? If the answer can have patient-related consequences, a review must take place before acting.
When should you consult your data protection officer?
Always involve them when:
- You set up a Claude business plan (Team/Enterprise)
- You want to use Claude in a process where patient data could occur
- You update your practice's record of processing activities (ROPA)
- You need a DPIA (data protection impact assessment)
- You are unsure whether a specific use is permissible
Usually not necessary when:
- You work with Claude privately (vacation planning, letter to the tax advisor)
- You do practice tasks without personal reference (marketing, job posting, answering machine script)
- You do general continuing education without reference to concrete cases
Family practices are often not required to appoint a DPO (that depends on staff numbers and data processing), but they carry the responsibility themselves. In that case: get external advice before using Claude in patient-related workflows. There are specialized data protection consultants for medical practices, often referred through the medical associations, KV, or professional bodies.
If you internalize the pre-use checklist, work on the right plan, recognize hallucinations, and preserve your medical responsibility, you can use Claude safely and productively. Data protection is not an obstacle — it's a routine you'll quickly internalize.
Ready to go in 10 minutes.
You need three things: a browser, an email address, five minutes of undisturbed time. This page walks you linearly from login to your first productive result. No technical prior knowledge required — just a little patience with yourself the first time.
"I've just poured a glass of wine. I'm opening Claude. I haven't done anything yet. Tell me exactly what to click — and what to type, so that after 10 minutes I know whether this is worth anything to me."
Create an account — like an email address, but in 90 seconds.
You do this once. After that, you just log in next time — exactly like on any other website.
Open claude.ai in your browser
Type it into the address bar of Chrome, Safari, Firefox, or Edge. You'll land on a start page with two buttons: "Sign up" and "Log in". If it's your first time there, click Sign up.
Sign up — easiest with Google
You can register with any email address. If you have a Google account (Gmail, Google Workspace), click "Sign in with Google" — one click, done. Alternatively, the classic way with an email address plus a password of your choice.
Confirm your phone number
Claude asks once for your mobile number to prevent spam accounts. You get a 6-digit code via SMS, enter it — that's it. Your number is not used for anything else.
Provide date of birth & name
A few mandatory fields for age verification and for how you're addressed. For the name, your first name is enough — you're talking to Claude, not a registry official. You can change this name at any time later.
You're in.
You now see a white chat window with a friendly greeting and a text input field at the bottom. That's all. This is where all the magic happens.
The Free plan is completely sufficient for what you'll do today. You can chat for several hours before you even hit a limit. Payment comes later — and only when you really need it. More on that below.
The interface — three areas, nothing more.
Don't get distracted by buttons you don't need today. These three areas are the ones you must know:
1. Text input field
At the bottom center. This is where you type or paste your question. Enter sends, Shift+Enter makes a line break. Everything happens here.
2. Left sidebar
Your history. Every chat you've had stays saved there. You can come back to it at any time. One click on "New chat" starts a fresh conversation.
3. Your initials (top right)
This takes you to the settings: change plan, profile details, privacy options. You'll only need this once today — we'll go there in step 4.
There are a few more icons — for Projects, Artifacts, models. Those come later. Today you only use the text input field. Really.
Writing your first prompt.
Let's make it sporting. Three tasks that will give you a real aha moment in the next 5 minutes. Simply copy each prompt, paste it into the input field, press Enter — and see what happens.
Test 1 — Translate a patient question into plain language
Test 2 — Draft an email
Test 3 — Something private: restaurant suggestions
You've just seen in three areas — medical explanation, business communication, private life — what Claude can do for you. Now let's set up a few more settings so this keeps working sensibly for you long-term.
Three settings you set once today.
Click on your initials at the top right, then on Settings. Three tabs matter:
Profile settings → "What should Claude call you?"
Enter your first name and briefly what you do. Example: "General practitioner with my own family practice in Berlin. I use Claude for patient communication, research, and private topics like family and travel." Claude will then tailor its answers to you better.
Privacy → training to "Off"
In the Privacy tab you'll find the option for whether Anthropic may use your chats to train new models. Set it to Off. That way your inputs won't end up in future training data. (On the Free plan this is important. On Team and Enterprise it's disabled by default anyway.)
Appearance → dark mode (optional)
Easy on the eyes, especially in the evening. A matter of personal taste. No impact on functionality — only on your eyes.
Real patient data still never belongs in Claude (at least not in the consumer plans Free/Pro/Max). Anthropic is a US provider, and no GDPR data processing agreement is available on the Free plan. For professional use involving patients, you need Claude for Work (Team/Enterprise) with a DPA or a different provider. Read the data protection chapter.
Which plan — and when?
Start on Free. Period. After two to four weeks of use you'll know for yourself whether to upgrade. Here is an honest, short overview (as of 2026):
Free — €0 / month
For whom: First-time users, occasional private use.
What you get: Claude in the chat, limited messages per day,
access to the current model, simple Projects.
How the limit feels: With intensive use you reach the daily
limit after 1–2 hours.
Pro — approx. €20 / month
For whom: Anyone who uses Claude regularly (several times a week).
What you get: ~5× more messages than Free, unlimited Projects,
extended knowledge base (RAG), longer conversations.
How the limit feels: Sufficient for most private users and occasional
practice tasks (without patient data).
Max — €100–200 / month
For whom: Power users who have Claude open all day.
What you get: 5×–20× higher limits than Pro, priority during
peak load.
How the limit feels: Practically none. Considerably oversized
for normal practice use.
Team / Enterprise
For whom: A practice with staff working with patient-related data.
What you get: Data processing agreement (DPA), no
training use of your data, staff management, shared Projects.
How the limit feels: The only path to GDPR-compliant use with
actual practice data.
Free for 2 weeks. Then upgrade to Pro when you notice that you're using Claude regularly. Only switch to Team before you concretely plan to use Claude with identifiable practice data — and even then only after consulting your data protection officer.
You're in. What now?
You've done the most important part. Three paths, depending on what appeals to you right now:
Understand the chat window
How to write really good prompts — the anatomy of a helpful prompt in 4 steps.
Go to the chat chapterUse cases in the practice
10 concrete scenarios — referral letter, patient education, research, marketing — as ready-made templates.
Practice use casesPrivate use cases
Family, vacation, gifts, letters to authorities, hobbies — make the invisible work of everyday life more efficient.
Private use cases