How Doctors Can Build a Personal AI Assistant Using Claude — Full Step-by-Step Guide

What Is Claude and Why Should Doctors Care?

If you have heard of ChatGPT, you already have the general idea. Claude is an AI assistant built by Anthropic — a company founded by former OpenAI researchers who left specifically to build safer, more reliable AI. Think of Claude as the colleague who actually reads the entire paper before discussing it, not just the abstract.

Here is why Claude stands out for medical professionals specifically:

It is genuinely good at medical reasoning. Claude scored above 85% on USMLE-style medical licensing exam questions. That does not make it a doctor — but it makes it a remarkably capable research assistant, documentation tool, and thinking partner for someone who is already a doctor.

It can read massive documents in one go. Claude has a 200,000 token context window. In plain terms, that means you can paste an entire 80-page clinical practice guideline, a full patient case file (anonymized, of course), or five research papers — all at once — and Claude will read and reference every word. No other mainstream AI tool matches this capacity. For a physician drowning in literature, this is transformative.

On January 12, 2026, Anthropic launched Claude for Healthcare with HealthEx connectors — purpose-built integrations designed for clinical workflows. This was not a marketing rebrand. It included new safety guardrails for medical contexts, improved citation handling, and structured output formats that map to clinical documentation standards.

The key advantage over other AI tools is that Claude was built with a focus on being honest about uncertainty. When it does not know something, it says so — instead of confidently generating wrong information. In medicine, that difference between "I think the answer is X" and "I am not sure, here is what I found" is the difference between useful and dangerous.

Let me be direct: Claude is not going to replace your clinical judgment. It cannot examine a patient. It cannot feel an abdomen. It cannot read the room when a patient is not telling you the full story. What it can do is handle the mountain of documentation, literature review, patient communication, and administrative work that eats up 40-60% of your day — so you can focus on the parts of medicine that actually require a human doctor.

Step 1: Set Up Your Claude Account

This takes about three minutes.

1. Go to claude.ai in your browser
2. Click "Sign Up" — use your email address (your personal email is fine; your hospital email works too)
3. Verify your email and set a password
4. You are in

The free tier lets you send messages and test things out. But if you are serious about building a proper medical assistant, you want Claude Pro at $20/month. Here is why:

• Unlimited messages — the free tier has daily limits that you will hit within an hour of real use
• Access to Projects — this is the feature that turns Claude from a chat tool into a persistent, customized assistant (more on this in Step 2)
• Priority access — no waiting in queues during peak hours
• Longer context windows — process larger documents without truncation

To upgrade, click your profile icon in the bottom-left corner, select "Upgrade to Pro", and enter your payment details. Standard credit/debit card, nothing complicated.

The Interface Basics

Once you are logged in, the interface is straightforward:

• Left sidebar — your conversation history, Projects, and settings
• Main chat area — where you type messages and read Claude's responses
• Attachment button — click the paperclip icon to upload PDFs, images, or documents directly into a conversation
• Model selector — at the top, you can choose between different Claude models. For medical work, always use the latest and most capable model available

Spend five minutes just chatting. Ask it a medical question you already know the answer to. See how it responds. Notice how it cites its reasoning, flags uncertainty, and structures its answers. This baseline familiarity matters before you start customizing.

Step 2: Create a Medical Project

This is where things get interesting. A Project in Claude is not just a folder — it is an isolated workspace where every conversation shares the same knowledge base and custom instructions. Think of it like giving Claude a specialty fellowship. Every time you start a new chat within this Project, Claude already knows who you are, what you need, and how you like things formatted.

Here is how to set it up:

1. In the left sidebar, click "Projects"
2. Click "New Project"
3. Name it something descriptive — for example: "Dr. Rahman — Clinical Assistant" or "Cardiology Practice AI"
4. Add a brief description if you want (optional, but helps if you create multiple projects later)
5. Click "Create"

You now have a dedicated workspace. Every chat you start inside this Project will have access to whatever custom instructions and knowledge base documents you add — which we are about to do.

Why does this matter? Without a Project, every new conversation with Claude starts from zero. You would need to re-explain your specialty, your preferences, your formatting requirements every single time. With a Project, you explain it once, and Claude remembers across every future conversation within that workspace. It is the difference between training a new intern every morning and having a reliable assistant who already knows how you work.

Step 3: Write Your Custom Instructions (The Secret Weapon)

This is the single most impactful thing you will do. Custom instructions tell Claude who you are, how to behave, and what rules to follow. Most people skip this step and wonder why their AI assistant gives generic, unhelpful responses. Do not be most people.

Inside your Project, click the gear icon or look for the "Project Instructions" area (usually in the right panel or project settings). This is where you paste a block of text that Claude reads at the start of every conversation.

Here is a complete template you can copy and customize:

You are a medical AI assistant for Dr. [Name], a [Specialty] physician at [Hospital/Clinic].

## Your Role
- You assist with clinical documentation, medical literature review, patient communication drafts, and medical education
- You are NOT a diagnostic tool — all clinical decisions are made by Dr. [Name]
- You always cite medical sources when providing clinical information
- You flag when information might be outdated or when guidelines vary by country

## Specialty Context
- Primary specialty: [e.g., Cardiology]
- Common conditions treated: [list]
- Preferred guidelines: [e.g., AHA/ACC, ESC, NICE, WHO]
- Hospital/clinic setting: [private practice / hospital / academic]

## Communication Style
- Use medical terminology when I ask clinical questions
- Use plain language (6th grade reading level) when I ask you to draft patient communication
- Be concise — I am a busy clinician
- When uncertain, say so explicitly rather than guessing

## Output Preferences
- Clinical notes: Use SOAP format unless I specify otherwise
- Medication lists: Include generic name, dose, route, frequency
- Literature summaries: Include study name, year, sample size, key findings, limitations
- Patient education: Use bullet points, short sentences, no jargon

## Safety Rules
- Never provide definitive diagnoses — frame as "considerations" or "differential"
- Always remind me to verify drug interactions with a pharmacist/database
- Flag any red-flag symptoms in patient scenarios
- Remind me this is an AI tool, not a replacement for clinical judgment

Let me walk through why each section matters.

Your Role sets boundaries. Without this, Claude might try to act like a diagnostic engine or give medical advice it has no business giving. By explicitly stating it is a documentation and research tool, you keep it in its lane.

Specialty Context is where the magic happens. A cardiologist and a dermatologist need fundamentally different things from an AI assistant. By specifying your specialty, common conditions, and preferred guidelines, Claude tailors every response to your clinical world. Instead of generic medicine, you get cardiology-specific (or whatever your field is) responses.

Communication Style handles the dual nature of medical work. When you are thinking through a differential, you want Claude speaking in medical terms. When you need a patient handout about their new metformin prescription, you want it written so a nervous patient can understand it at 2 AM. This instruction set handles both modes.

Output Preferences eliminate the back-and-forth of "no, format it this way." Tell Claude your SOAP format preference once, and every clinical note follows it. Tell it how you want medication lists structured, and it is consistent every time.

Safety Rules are non-negotiable. These prevent Claude from overstepping. You do not want an AI that says "the patient has pneumonia" — you want one that says "based on the described findings, considerations include community-acquired pneumonia, and the following differentials should be evaluated." That framing keeps AI in its proper role as a support tool.

Paste the template above into your Project Instructions. Replace every [bracket] with your actual information. Be specific — the more detail you provide about your specialty and preferences, the better Claude performs from the very first message.

Step 4: Build Your Knowledge Base

Custom instructions tell Claude how to think. The Knowledge Base tells Claude what to know. Together, they turn a general-purpose AI into something that feels like it did a fellowship in your specialty.

Inside your Project, look for the Knowledge Base panel (usually on the right side). This is where you upload documents that Claude reads automatically in every conversation within this Project.

What to Upload

Think about the documents you reference repeatedly in your daily practice:

• Clinical practice guidelines — your specialty's primary guidelines (PDF). For a cardiologist, that might be the latest AHA/ACC guidelines on heart failure management. For an endocrinologist, the ADA Standards of Care.
• Hospital formulary or preferred drug lists — so Claude recommends medications that are actually available in your pharmacy
• Your clinic's standard operating procedures — admission protocols, discharge checklists, referral pathways
• Template letters — referral letters, medical certificates, insurance prior authorization formats you use regularly
• Your preferred SOAP note format — upload an example of a SOAP note you wrote and liked, so Claude mirrors your style
• Patient education materials — handouts you frequently give patients, so Claude can generate similar ones for new conditions
• Local clinical protocols — if your hospital has specific antibiotic stewardship guidelines, sepsis bundles, or clinical pathways, upload those

Supported file types: PDF, DOCX, CSV, TXT, HTML — up to 30MB per file. PDFs of clinical guidelines work particularly well because Claude can reference specific sections and page numbers.

Here is what happens after you upload: every time you start a new chat within this Project, Claude has already read all of those documents. So when you say "draft a referral letter for a patient with uncontrolled hypertension," Claude does not just generate a generic letter — it follows your clinic's referral template format, references the hypertension guidelines you uploaded, and uses the terminology style from your existing letters.

That is the difference between a generic AI and a personal assistant that knows your practice.

Step 5: Set Up Custom Skills (For Claude Code Users)

If you are using Claude Code — Anthropic's desktop and command-line interface — you have access to a powerful feature called SKILL.md files. These are markdown files that extend Claude with specialized, repeatable capabilities.

Think of a SKILL.md file like a detailed SOP for Claude. Instead of re-explaining how to write a SOAP note every time, you create a skill file once, and Claude activates it automatically when relevant.

How a SKILL.md File Works

Each SKILL.md file has two parts:

1. YAML frontmatter — a small header section that tells Claude the skill's name and when to use it
2. Markdown body — detailed instructions for how to execute the skill

Here is a complete SOAP Note Generator skill:

---
name: soap-note-generator
description: Generate SOAP notes from patient encounter descriptions. Use when the doctor describes a patient visit and wants structured clinical documentation.
---

# SOAP Note Generator

When given a patient encounter description, generate a structured SOAP note.

## Format
**Subjective:** Patient's chief complaint, history of present illness, review of systems
**Objective:** Vital signs, physical exam findings, lab/imaging results mentioned
**Assessment:** Clinical impression, differential diagnosis considerations
**Plan:** Treatment plan, medications, follow-up, referrals, patient education

## Rules
- Use standard medical abbreviations
- Flag any mentioned red-flag symptoms with a warning marker
- Include ICD-10 code suggestions where applicable
- Always end with "Note: This is AI-generated and must be reviewed by the attending physician"

Save this as soap-note-generator.SKILL.md in your project directory. Claude Code detects it automatically.

Patient Education Generator Skill

Here is a second skill for creating patient handouts:

---
name: patient-education-generator
description: Generate plain-language patient education handouts. Use when the doctor asks for patient-friendly explanations of conditions, medications, or procedures.
---

# Patient Education Generator

Create patient education materials that are clear, accurate, and accessible.

## Output Format
- Title: Condition/topic name in plain language
- What It Is: 2-3 sentence explanation at 6th grade reading level
- Key Facts: 4-6 bullet points patients need to know
- What You Can Do: Practical self-care steps
- When to Call Your Doctor: Red-flag symptoms in plain language
- Questions to Ask at Your Next Visit: 3-4 suggested questions

## Rules
- No medical jargon — explain terms if they must be used
- Use short sentences (under 20 words each)
- Include specific numbers where helpful (e.g., "call if fever above 101.3F")
- Culturally sensitive — avoid assumptions about diet, lifestyle, or family structure
- End with: "This handout was created to support your understanding. It does not replace advice from your doctor."

Literature Review Assistant Skill

And a third skill for structured research:

---
name: literature-review-assistant
description: Help structure clinical literature searches and summarize research papers. Use when the doctor asks about evidence for a treatment, wants a paper summarized, or needs help formulating a research question.
---

# Literature Review Assistant

Help the doctor with evidence-based medicine tasks.

## When Given a Clinical Question
1. Reformat it as a PICO question (Population, Intervention, Comparison, Outcome)
2. Suggest MeSH terms and search strategies for PubMed
3. Identify the type of evidence needed (RCT, meta-analysis, cohort, etc.)
4. Note any relevant Cochrane reviews on the topic

## When Given a Research Paper
Summarize using this structure:
- **Citation:** Authors, title, journal, year
- **Study Design:** Type (RCT, cohort, case-control, etc.) and level of evidence
- **Population:** Sample size, demographics, inclusion/exclusion criteria
- **Intervention vs. Control:** What was compared
- **Key Findings:** Primary and secondary outcomes with numbers
- **Limitations:** Study weaknesses, bias risk, generalizability concerns
- **Clinical Relevance:** What this means for practice, in one paragraph

## Rules
- Always note the level of evidence
- Flag if the study is industry-funded
- Note if findings contradict current guidelines
- Distinguish between statistical significance and clinical significance

These three skills alone — SOAP notes, patient education, and literature review — cover a significant chunk of a doctor's daily non-clinical workload. You can create additional skills for discharge summaries, referral letters, conference prep, or anything else you do repeatedly.

Step 6: CLAUDE.md — Your AI Memory File

For doctors using Claude Code (the desktop/CLI version), there is one more configuration file that ties everything together: CLAUDE.md.

This file lives in your project root directory. Claude reads it at the very start of every session — before you type a single message. It functions as persistent memory. While custom instructions in the web interface reset if you start a new project, CLAUDE.md stays with your local project permanently.

Here is a practical CLAUDE.md template for a medical practice:

# Dr. [Name] Clinical Assistant

## Practice Details
- Specialty: [Cardiology]
- Location: [City, Country]
- EMR System: [Epic / mediONE / paper-based]

## Common Tasks
- SOAP note generation from voice/text descriptions
- Patient education material creation
- Medical literature review and summarization
- Referral letter drafting
- ICD-10 code lookup assistance

## Important Rules
- NEVER include real patient identifiers (name, DOB, ID numbers)
- All clinical notes must include disclaimer about AI-assisted generation
- Follow [AHA/ACC] guidelines as primary reference
- Drug dosing should reference [BNF / UpToDate / local formulary]
- Keep all output HIPAA/data privacy compliant

Save this as CLAUDE.md in the root of your project folder. Claude Code reads it automatically every time you launch a session.

The beauty of CLAUDE.md is its simplicity. It is just a markdown text file. You can edit it in any text editor — Notepad, VS Code, even the Notes app on your phone if you sync your files. Every time you learn something about how you want Claude to behave, add it to CLAUDE.md. Over weeks and months, it becomes an increasingly refined instruction set that makes your AI assistant better with every update.

10 Practical Prompts Every Doctor Should Save

Theory is great. Practical prompts you can copy, paste, and use right now are better. Here are ten prompts that cover the most common physician workflows. Each one is tested and ready to use with Claude.

1. SOAP Note from Description

Write a SOAP note based on the following patient encounter. Use standard medical abbreviations. Include ICD-10 code suggestions. Flag any red-flag symptoms.

Patient encounter:
[Paste your encounter description here — remember, use anonymized/fictional patient details only]

2. Differential Diagnosis Brainstorm

I need to think through a differential diagnosis. Here is the clinical picture:

[Describe the presentation — age, sex, chief complaint, relevant history, exam findings, labs]

Generate a ranked differential diagnosis with:
- Most likely diagnoses (with reasoning)
- Must-not-miss diagnoses (even if less likely)
- Suggested next investigations for each

Remind me that this is an AI-generated list for consideration only, not a clinical diagnosis.

3. Drug Interaction Check Prompt

I am considering the following medication combination for a patient. Please review for potential interactions, contraindications, and dosing considerations:

Medications:
1. [Drug name, dose, route, frequency]
2. [Drug name, dose, route, frequency]
3. [Drug name, dose, route, frequency]

Patient context: [age, sex, relevant conditions like renal/hepatic impairment]

Note: I will verify this against a pharmacist and an up-to-date drug interaction database. This is for initial review only.

4. Patient Education Handout

Create a patient education handout about [condition name] for a patient who was just diagnosed. Write at a 6th grade reading level. No medical jargon. Include:
- What the condition is (2-3 sentences)
- What causes it
- What the patient can do (lifestyle, diet, activity)
- What their medication does and common side effects
- When to call the doctor (red flag symptoms)
- 3 questions they should ask at their next visit

5. Referral Letter Template

Draft a referral letter from me (Dr. [Your Name], [Your Specialty]) to a [Referred Specialty] colleague. The letter should be professional, concise, and include:

Patient: [Anonymized description — age, sex only]
Reason for referral: [Clinical question]
Relevant history: [Brief summary]
Current medications: [List]
What I have tried: [Previous management]
What I am asking for: [Specific request — opinion, procedure, co-management]

Format it as a standard medical referral letter.

6. Medical Certificate Draft

Draft a medical certificate for the following:

Purpose: [Fitness to work / Fitness to travel / Sick leave / School absence]
Duration: [Dates]
Patient: [Use fictional details — age, sex only]
Relevant details: [What the certificate needs to state]

Include appropriate disclaimers. Keep the language professional and legally appropriate for [country].

7. Research Paper Summary

Summarize the following research paper. Structure the summary as:
- Citation (authors, journal, year)
- Study design and level of evidence
- Population (sample size, demographics)
- Intervention vs. control
- Key findings (with numbers)
- Limitations
- Clinical relevance (what this means for my practice)

[Paste the paper abstract or full text here]

8. Treatment Plan Comparison

Compare the following treatment approaches for [condition] in a [patient description — age, sex, comorbidities]:

Option A: [Treatment 1]
Option B: [Treatment 2]
Option C: [Treatment 3]

For each option, outline:
- Mechanism of action
- Expected efficacy (cite trials if possible)
- Side effect profile
- Cost considerations
- Monitoring requirements
- Contraindications relevant to this patient

Present as a comparison table followed by a brief clinical discussion.

9. Discharge Summary from Notes

Generate a discharge summary from the following admission notes. Follow standard discharge summary format:

- Patient demographics (anonymized)
- Admission diagnosis
- Hospital course
- Procedures performed
- Discharge diagnosis
- Discharge medications (with changes highlighted)
- Follow-up plan
- Patient education provided
- Red flags for re-presentation

Admission notes:
[Paste your notes here — anonymized]

10. Conference Presentation Outline

I need to prepare a [case presentation / journal club / grand rounds / CME lecture] on [topic].

Audience: [Medical students / Residents / Attendings / Mixed]
Duration: [Time limit]
Format: [Slide presentation / Oral / Poster]

Create an outline with:
- Key learning objectives (3-5)
- Slide-by-slide structure with suggested content
- Key references to cite
- Discussion points / questions for the audience
- Take-home messages

Make it clinically relevant and evidence-based.

What Claude Cannot Do (Important Limitations)

I want to be straightforward about this because overselling AI tools to doctors is irresponsible. Claude is powerful, but it has real limitations that you need to understand before you integrate it into any aspect of your practice.

It Is Not a Diagnostic Tool

Claude can help you brainstorm differentials. It can remind you of rare conditions you might have forgotten from medical school. But it cannot examine a patient, interpret the subtle nuances of a presentation, or exercise the clinical judgment you have built over years of training and practice. Using it as a diagnostic tool would be like using a textbook as a doctor — the information might be there, but the judgment is not.

It Is Not Connected to Real-Time Medical Databases

Claude does not have live access to UpToDate, PubMed, Lexicomp, or any drug interaction database. When it tells you about a drug interaction, it is drawing from its training data — which may not include the latest safety alerts or newly published interactions. Always verify drug information with a current, authoritative database. This is not optional. It is mandatory.

It Can Hallucinate Medical Facts

This is the biggest risk. Claude can generate information that sounds completely authoritative — correct formatting, confident language, plausible-sounding citations — that is factually wrong. It can cite studies that do not exist. It can state drug doses that are incorrect. It can describe guidelines that were never published. This is not a bug that will be fixed — it is a fundamental characteristic of how large language models work. Every piece of clinical information Claude provides must be independently verified.

It Is Not HIPAA Compliant by Default

The consumer version of Claude at claude.ai is not covered by a Business Associate Agreement (BAA). That means uploading real Protected Health Information violates HIPAA (in the US) and equivalent privacy regulations in other countries. Anthropic does state that conversation data is not used for model training, but that is not the same as HIPAA compliance. For clinical use involving real patient data, you need enterprise-grade solutions with proper BAAs — talk to your hospital's IT and compliance teams.

Training Data Has a Cutoff

Claude's medical knowledge comes from its training data, which has a cutoff date. Guidelines updated after that date may not be reflected. New drugs approved recently may not be in its knowledge base. New safety warnings may not be known. This is why uploading your specialty's latest guidelines to the Knowledge Base is so important — it supplements Claude's built-in knowledge with current information.

It Cannot Access Your EMR

Claude cannot log into Epic, mediONE, Cerner, or any electronic medical records system. It cannot pull up patient charts, lab results, or imaging reports on its own. You have to manually provide (anonymized) information for it to work with. The HealthEx connectors announced in January 2026 are a step toward EMR integration, but widespread clinical EMR integration is still evolving and requires institutional-level decisions about data governance.