Prompt Engineering for Healthcare

Build a clinical AI assistant in Claude with safety guardrails.

Introduction

⚡️ 30 Second Summary

Over 76% of physicians already use AI chatbots for clinical tasks, but most use generic prompts that produce generic, unreliable results.

In this project, you will set up a Claude Project as a persistent clinical AI assistant, learning four core prompt engineering techniques and applying them to real healthcare use cases with safety guardrails.

What You'll Build

You'll configure a Claude Project with specialty-specific custom instructions that acts as a reusable clinical AI assistant for drug interactions, patient education, and differential diagnosis.

Your Claude Project stores persistent instructions that power every clinical prompt — role prompting, audience adaptation, structured output, and a safety layer — so each technique builds on the last.

By the end of this project, you'll have:

  • 💊 Drug interaction prompts using role prompting to simulate a clinical pharmacist.
  • 📖 Patient education materials rewritten at 5th-grade, 9th-grade, and clinical reading levels.
  • 🩺 Differential diagnosis outputs with structured confidence levels, red flags, and next steps.
  • 💎 Secret Mission: A clinical safety layer with citation requirements, uncertainty flagging, and diagnosis refusal.

Want a complete demo of how to do this project, from start to finish? Check out our 🎬 walkthrough with Maya

Do I need to pay to do this project?

This project is completely free. You only need a Claude free-tier account at claude.ai, which gives you access to Claude Projects at no cost.

Not sure if this project is right for you? Check if it matches your goals

If you're up for a bit of a challenge, quiz yourself on the key concepts up ahead in this project.

Get Set Up

Most physicians use AI chatbots without any configuration, which means every conversation starts from scratch with no clinical context. To build a reliable clinical assistant, you need a persistent workspace that remembers your specialty and enforces safety guardrails from the start.

In this step, you will create a Claude Project that carries your custom instructions into every conversation, so you never have to repeat your clinical context again.

In this step, get ready to:

  • Create a free Claude account at claude.ai.
  • Create a new Claude Project called "Clinical Assistant."
  • Add base Project instructions with your medical specialty and role context.

This project uses synthetic data only.

Everything in this project uses sample and synthetic data. In production with real patient data, you would need Claude Enterprise with a Business Associate Agreement (BAA). This project focuses on non-PHI use cases physicians can start using immediately.

Create Your Claude Account

Claude is an AI assistant built by Anthropic. The free tier uses Claude Sonnet 4.6, which is sufficient for all tasks in this project.

✔️ I already have an account

  • Navigate to claude.ai and log in with your existing account.

ⓧ I don't have an account yet

  • Navigate to claude.ai.
  • Click Sign up.
  • Choose your sign-up method (Google, email, or SSO).
  • Verify your email if prompted.

Why Claude?

Claude supports a feature called Claude Projects that lets you set persistent instructions and upload reference documents. Other AI chatbots require you to repeat context every time, but a Claude Project carries your instructions forward automatically.

Create the Clinical Assistant Project

  • Click Projects in the left sidebar.
  • Click New Project (or Create Project).
  • Name the project Clinical Assistant.
  • In the What are you trying to achieve? field, type a short description like Build a clinical AI assistant with safety guardrails, drug interaction checking, patient education, and differential diagnosis support.

What is a Claude Project?

Claude Projects are persistent workspaces with custom instructions, uploaded documents, and saved conversation history. The free tier allows up to 5 projects. Every conversation inside the project automatically uses your instructions, so you do not need to repeat context each time.

Add Base Project Instructions

Now you will add custom instructions that tell Claude how to behave in every conversation within this Project. These instructions define your assistant's role, safety guardrails, and specialty context.

  • Click Set project instructions (or the gear icon in the project).

What are Project instructions?

Project instructions act as a system prompt that runs before every conversation. Think of them as standing orders for your AI assistant. Anything you put here applies automatically, so you do not need to repeat it each time you chat.

  • Paste the following instructions:
You are a clinical AI assistant supporting a physician. Your role is to help
with clinical reasoning, drug information lookup, patient education, and
diagnostic brainstorming.

IMPORTANT LIMITATIONS:
- You are NOT a replacement for clinical judgment.
- All outputs are for educational/brainstorming purposes only.
- Never use real patient data in this chat.
- Always recommend verification with authoritative sources (UpToDate,
  Lexicomp, clinical guidelines).

My specialty: [[YOUR_SPECIALTY="Family Medicine"]]
My practice setting: [[YOUR_PRACTICE_SETTING="outpatient clinic"]]
  • Replace the placeholder values with your own specialty and practice setting.
  • Click Save.

Why add safety disclaimers to the instructions?

By embedding disclaimers directly in the project instructions, every response Claude generates will include appropriate caveats. This builds a habit of treating AI output as a brainstorming tool, not a clinical authority.

Your clinical assistant is configured and ready. Next, you will put it to work checking drug interactions using a powerful prompt engineering technique called role prompting.

Check Drug Interactions

Your Clinical Assistant Project is set up with base instructions. Now it is time to give it its first real clinical task: checking drug interactions.

This is one of the most common physician use cases for AI, but generic prompts produce generic and sometimes dangerously incomplete answers. Research from a 2024 JAMA Network Open study shows LLMs have roughly 78% accuracy on drug interactions compared to 94% for dedicated tools like Lexicomp. Better prompts close that gap significantly.

In this step, get ready to:

  • Use role prompting to make Claude respond as a clinical pharmacist.
  • Test drug interaction checks with sample medication pairs.
  • Iterate on your prompt to improve specificity and clinical relevance.

Claude is not a substitute for clinical databases.

Tools like Lexicomp, UpToDate, and Medscape remain the gold standard for drug interaction checking. Claude has a specificity of roughly 0.64 for drug interactions, meaning it can produce false positives for interactions that do not exist. Use Claude as a brainstorming and triage tool, not a definitive clinical reference.

Apply Role Prompting

Role prompting means telling Claude to adopt a specific expert persona. For drug interactions, a clinical pharmacist role produces more thorough, mechanism-aware responses than a generic assistant.

  • Navigate to the Project instructions in your Clinical Assistant project.
  • Click the gear icon or Project instructions to open the settings.
  • Add the following role prompting instructions to your existing Project instructions:
When I ask about drug interactions, respond as a clinical pharmacist with
10+ years of experience. For each interaction:
1. Severity level (major/moderate/minor)
2. Mechanism of interaction
3. Clinical significance
4. Management recommendation
5. Source/guideline reference if known

Why a clinical pharmacist role?

Pharmacists specialize in drug mechanisms and interactions. By telling Claude to respond as one, you get structured, mechanism-aware answers instead of vague safety warnings. This is what makes role prompting powerful: the persona shapes the depth and format of the response.

  • Save your updated Project instructions.

Test Drug Interaction Checks

Now test your role prompt with three sample medication pairs that cover different interaction severity levels.

  • In your Project, start a new conversation by typing the following prompt in the text box:
What is the drug interaction between Warfarin and Aspirin?
  • Press Enter and review Claude's response.

Why this example?

Warfarin and Aspirin is a well-known major interaction. Claude should identify the increased bleeding risk, the mechanism involving anticoagulant and antiplatelet effects, and recommend monitoring or avoidance. This is a good first test because the answer is well-documented and easy to verify.

  • Now test a common but generally safe combination. Ask about Metformin and Lisinopril.
What is the drug interaction between Metformin and Lisinopril?

Claude should indicate this is a generally safe combination commonly prescribed together for patients with diabetes and hypertension.

  • Finally, test a more nuanced interaction. Ask about Sertraline and Tramadol.
What is the drug interaction between Sertraline and Tramadol?

This pair carries a serotonin syndrome risk. Claude should flag the severity, explain the serotonergic mechanism, and recommend clinical monitoring or alternative medications.

What is serotonin syndrome?

Serotonin syndrome is a potentially life-threatening condition caused by excessive serotonergic activity. It can occur when two drugs that increase serotonin levels are taken together. Symptoms range from mild (tremor, diarrhea) to severe (high fever, seizures, muscle rigidity).

Iterate on Prompt Precision

Your role prompt works, but there is room to tighten Claude's output. Look at Claude's Warfarin-Aspirin response and check: did it clearly say whether the interaction is theoretical or clinically documented? Did it mention when the interaction kicks in (immediately or over days)? If either is missing, the next set of constraints will fix that.

  • Go back to your Project instructions and add these constraints below your existing role prompt:
Additional drug interaction rules:
- Always state if this is a theoretical or clinically documented interaction.
- Include onset timing (immediate vs. delayed) when relevant.
- If you are uncertain about an interaction, say so explicitly.
  • Save your updated instructions and test the same drug pairs again.

Notice how the added constraints produce more specific, actionable responses. Claude now distinguishes between theoretical and documented interactions and flags its own uncertainty.

  • Try a polypharmacy scenario with 5 or more medications to see how Claude handles complex lists:
Check for interactions between the following medications:
Lisinopril, Metformin, Atorvastatin, Amlodipine, and Sertraline
  • Try an obscure or lesser-known drug combination to see if Claude admits uncertainty versus fabricating an interaction.

Claude confidently listed an interaction that does not exist?

This is expected. LLMs have a specificity of roughly 0.64 for drug interactions, meaning they produce false positives. This is exactly why we added the "if you are uncertain, say so explicitly" constraint. If Claude still fabricates, try adding: "Do not invent interactions. If you cannot find a documented interaction, state that clearly." How can I reduce Claude's hallucinations for drug interactions?

Your drug interaction prompts are dialed in. Next, you will tackle a completely different challenge: taking complex medical content and making it understandable for patients at different reading levels.

Simplify Medical Content for Patients

Your drug interaction checker is working well with role prompting. Now you have a different problem: patient communication. What happens when you need to explain a complex condition to a patient with limited medical vocabulary? What if you need the same explanation at three different reading levels? Health literacy studies show that nearly 36% of US adults have limited health literacy, making this a critical skill.

Claude can rewrite medical content at specific reading levels, but only if you prompt it correctly. In this step, you will use Claude Styles to create reusable writing presets for different reading levels, then apply them to clinical text using audience adaptation prompting.

In this step, get ready to:

  • Provide Claude with sample clinical text to simplify.
  • Use audience adaptation to rewrite content at 5th-grade, 9th-grade, and clinical levels.
  • Create reusable Claude Styles for each reading level.

Provide Clinical Source Text

Before you can simplify anything, you need a piece of clinical text to work with. You will use a textbook-style description of Type 2 Diabetes management as your sample.

  • Open your Clinical Assistant project in Claude.ai.
  • Start a new conversation by clicking on the text box at the bottom of the project page.
  • Paste the following clinical paragraph into your chat:
Type 2 Diabetes Mellitus (T2DM) management requires a multifaceted approach including glycemic control through pharmacotherapy (metformin as first-line, with potential addition of SGLT2 inhibitors, GLP-1 receptor agonists, or insulin), lifestyle modifications (medical nutrition therapy targeting 5-7% weight reduction, 150 min/week moderate-intensity physical activity), and regular monitoring (HbA1c every 3-6 months, annual comprehensive metabolic panel, dilated eye exam, and foot examination). Patients with comorbid hypertension should target BP <130/80 mmHg, and those with dyslipidemia should be initiated on high-intensity statin therapy for cardiovascular risk reduction.
  • Ask Claude to explain this text without any specific audience instructions. For example, type:
Can you explain this medical text in simpler terms?

Claude responds with a general simplification. This is your baseline to compare against the targeted rewrites you will create next.

Why establish a baseline first?

Without specific audience instructions, Claude guesses who it is writing for. The result is usually somewhere between clinical and casual. By seeing this default first, you can appreciate how much difference targeted audience adaptation makes.

Apply Audience Adaptation Prompting

Now you will use audience adaptation prompting to rewrite the same clinical text at three different reading levels. This technique works by specifying the target audience and reading level explicitly in your prompt.

  • In the same conversation, paste the following prompt to generate a 5th-grade version:
Rewrite the following medical text for a patient with a 5th grade reading level. Requirements:
- No medical jargon (or define it in simple terms if unavoidable)
- Short sentences (max 15 words per sentence)
- Use analogies to explain complex concepts
- Include a "What to do" action section
- Maintain medical accuracy — do not oversimplify to the point of being wrong

Type 2 Diabetes Mellitus (T2DM) management requires a multifaceted approach including glycemic control through pharmacotherapy (metformin as first-line, with potential addition of SGLT2 inhibitors, GLP-1 receptor agonists, or insulin), lifestyle modifications (medical nutrition therapy targeting 5-7% weight reduction, 150 min/week moderate-intensity physical activity), and regular monitoring (HbA1c every 3-6 months, annual comprehensive metabolic panel, dilated eye exam, and foot examination). Patients with comorbid hypertension should target BP <130/80 mmHg, and those with dyslipidemia should be initiated on high-intensity statin therapy for cardiovascular risk reduction.
  • Now generate the 9th-grade version. Paste the same prompt but change the reading level:
Rewrite the following medical text for a patient with a 9th grade reading level. Requirements:
- No medical jargon (or define it in simple terms if unavoidable)
- Minimize medical jargon (define terms when used)
- Clear, readable sentences
- Use analogies to explain complex concepts
- Include a "What to do" action section
- Maintain medical accuracy — do not oversimplify to the point of being wrong

Type 2 Diabetes Mellitus (T2DM) management requires a multifaceted approach including glycemic control through pharmacotherapy (metformin as first-line, with potential addition of SGLT2 inhibitors, GLP-1 receptor agonists, or insulin), lifestyle modifications (medical nutrition therapy targeting 5-7% weight reduction, 150 min/week moderate-intensity physical activity), and regular monitoring (HbA1c every 3-6 months, annual comprehensive metabolic panel, dilated eye exam, and foot examination). Patients with comorbid hypertension should target BP <130/80 mmHg, and those with dyslipidemia should be initiated on high-intensity statin therapy for cardiovascular risk reduction.
  • Finally, generate a clinical version for referring physicians:
Rewrite the following medical text as a concise clinical summary for a referring physician. Requirements:
- Use standard medical terminology and abbreviations
- Prioritize actionable clinical information
- Include specific targets, dosing considerations, and monitoring intervals
- Structure as a brief consultation note

Type 2 Diabetes Mellitus (T2DM) management requires a multifaceted approach including glycemic control through pharmacotherapy (metformin as first-line, with potential addition of SGLT2 inhibitors, GLP-1 receptor agonists, or insulin), lifestyle modifications (medical nutrition therapy targeting 5-7% weight reduction, 150 min/week moderate-intensity physical activity), and regular monitoring (HbA1c every 3-6 months, annual comprehensive metabolic panel, dilated eye exam, and foot examination). Patients with comorbid hypertension should target BP <130/80 mmHg, and those with dyslipidemia should be initiated on high-intensity statin therapy for cardiovascular risk reduction.
  • Scroll through the conversation to review all three versions.

Notice how the 5th-grade version uses analogies and short sentences, the 9th-grade version introduces some medical terms with definitions, and the clinical version is dense with abbreviations and specific targets.

What is audience adaptation prompting?

Audience adaptation is a prompt engineering technique where you explicitly tell the AI who the reader is and what reading level to target. The same information gets restructured, not just shortened. A 5th-grade version uses analogies and avoids jargon entirely, while a clinical version assumes deep domain knowledge.

Refine for Accuracy and Completeness

Simplified text can accidentally lose critical safety information. You need to check your outputs and add constraints to prevent this.

  • Review your 5th-grade version for missing safety information. Check whether medication names are still present and whether warning signs are mentioned.
  • Check whether the analogies are accurate or misleading.

A bad analogy can be worse than no analogy at all.

  • Add constraints to your prompt to improve safety. Paste the following refined prompt:
Rewrite the following medical text for a patient with a 5th grade reading level. Requirements:
- No medical jargon (or define it in simple terms if unavoidable)
- Short sentences (max 15 words per sentence)
- Use analogies to explain complex concepts
- Include a "What to do" action section
- Maintain medical accuracy — do not oversimplify to the point of being wrong
- Always include warning signs that require calling a doctor
- Never remove medication names, even in simplified versions
- Include a disclaimer that this is educational content, not medical advice

Type 2 Diabetes Mellitus (T2DM) management requires a multifaceted approach including glycemic control through pharmacotherapy (metformin as first-line, with potential addition of SGLT2 inhibitors, GLP-1 receptor agonists, or insulin), lifestyle modifications (medical nutrition therapy targeting 5-7% weight reduction, 150 min/week moderate-intensity physical activity), and regular monitoring (HbA1c every 3-6 months, annual comprehensive metabolic panel, dilated eye exam, and foot examination). Patients with comorbid hypertension should target BP <130/80 mmHg, and those with dyslipidemia should be initiated on high-intensity statin therapy for cardiovascular risk reduction.

The refined version now keeps medication names, includes warning signs, and adds a disclaimer.

Why keep medication names in simplified text?

Patients need to recognize their medication names on prescription labels and bottles. Removing medication names from simplified text creates a disconnect between what the patient reads and what they actually take. Simplify the explanation around the name, but keep the name itself.

  • Now that you have a refined prompt that works well, save it as a reusable Claude Style so you can apply it to any text without retyping the prompt every time.
  • Start a new conversation in your Clinical Assistant project by clicking on the text box at the bottom of the project page.
  • Click the + (plus icon) in the bottom left corner of the text box.
  • Click Use style.
  • Click Create & edit styles.
  • Click Create custom style.
  • Click Describe style instead.

Now you will tailor your style based on your needs.

  • Select Tailor to an audience.
  • In the style instructions, paste your refined audience adaptation prompt:

Give me a prompt to use

Rewrite any medical text for a patient with a 5th grade reading level. Requirements:
- No medical jargon (or define it in simple terms if unavoidable)
- Short sentences (max 15 words per sentence)
- Use analogies to explain complex concepts
- Include a "What to do" action section
- Maintain medical accuracy — do not oversimplify to the point of being wrong
- Always include warning signs that require calling a doctor
- Never remove medication names, even in simplified versions
- Include a disclaimer that this is educational content, not medical advice

I'll write my own

Use the refined prompt you built in this step as a template. Make sure it includes constraints for reading level, jargon rules, safety information, and a disclaimer.

  • Click Generate style.
  • To name the style, click on the text on the top of the panel and name it Patient Education - 5th Grade.

Why use Styles instead of Project instructions?

Claude Styles are designed for controlling how Claude writes. They are perfect for audience adaptation because you can switch between reading levels with one click instead of editing Project instructions each time. Styles also work across all your projects and conversations, so you can reuse your patient education styles anywhere.

Let's test your new Style.

  • Start a new conversation in your Clinical Assistant project by clicking on the text box at the bottom of the project page.
  • Enter the text to be explained in 5th Grade style.
Type 2 Diabetes Mellitus (T2DM) management requires a multifaceted approach including glycemic control through pharmacotherapy (metformin as first-line, with potential addition of SGLT2 inhibitors, GLP-1 receptor agonists, or insulin), lifestyle modifications (medical nutrition therapy targeting 5-7% weight reduction, 150 min/week moderate-intensity physical activity), and regular monitoring (HbA1c every 3-6 months, annual comprehensive metabolic panel, dilated eye exam, and foot examination). Patients with comorbid hypertension should target BP <130/80 mmHg, and those with dyslipidemia should be initiated on high-intensity statin therapy for cardiovascular risk reduction.
  • After entering the prompt, type /
  • Then select Use style, and... do you see the style you added?
  • Once you've selected the Patient Education - 5th Grade style from the style picker, press Enter and check out the output!

Skills vs. Styles — when to use which

Project instructions and Styles run in every conversation automatically. A Claude Skill is different: you activate it only when you need it. This matters because you do not always want Claude to explain things at a 5th-grade level. Sometimes you need a clinical summary, sometimes a patient handout. Skills let you pick the right behavior for the moment instead of having one mode running all the time.

You now have reusable Styles for patient education at different reading levels. Next, you will build a structured differential diagnosis assistant using Claude Skills.

Build a Differential Diagnosis Assistant

You have drug interaction checking and patient education Styles working. Now it is time for the most intellectually demanding use case: differential diagnosis support. What happens when you ask an AI to brainstorm a differential? Without structured prompting, Claude returns unorganized lists that are hard to act on clinically. How do you get ranked diagnoses with confidence levels, red flags, and next steps in a consistent format?

In this step, you will create a Claude Skill that generates structured differential diagnoses on demand. Unlike Styles (which change how Claude writes), Skills define a specific task that you can trigger whenever you need it.

In this step, get ready to:

  • Create a Claude Skill for structured differential diagnosis.
  • Test with a sample clinical presentation.
  • Iterate to include confidence levels, red flags, and next steps.

Create a Differential Diagnosis Skill

A Claude Skill is a saved prompt template that you can invoke on demand. Instead of pasting the same structured prompt every time, you create a Skill once and run it whenever you need a differential diagnosis.

Projects, Styles, and Skills on the free plan

Claude's free tier limits the number of Projects you can create and the number of messages you can send. If you are running into usage limits or want to create more Projects, Styles, and Skills, consider upgrading to Claude Pro for a higher usage allowance and unlimited Projects.

  • Start a new conversation in your Clinical Assistant project by clicking on the text box at the bottom of the project page.
  • Click the + (plus icon) in the bottom left corner of the text box.
  • Click Skills in the left sidebar.
  • Click Manage skills.
  • Click the + (plus icon) in the top corner of the Skills sidebar.
  • Click on Write skills instructions.
  • Name the skill Differential Diagnosis.
  • Under the Description, enter Generate differential diagnosis report with Rank, Cannot-Miss Diagnoses and Summary.
  • In the skill instructions, paste the following structured output template:
When I present a clinical case for differential diagnosis, respond in this
exact format:

## Differential Diagnosis

For each diagnosis (list 5-7, ranked by likelihood):

### [Rank]. [Diagnosis Name]
- **Likelihood:** [High / Moderate / Low]
- **Key supporting features:** [from the presented case]
- **Key features against:** [from the presented case]
- **Red flags to rule out:** [urgent findings that would confirm/exclude]
- **Recommended next steps:** [specific tests, imaging, or referrals]

## Cannot-Miss Diagnoses
[List any dangerous diagnoses that must be ruled out regardless of likelihood]

## Summary
[2-3 sentence clinical reasoning summary]
  • Click Create.

Why use a Skill instead of Project instructions?

A Skill is designed for specific, repeatable tasks. The differential diagnosis format only applies when you are analyzing a clinical case, not in every conversation. By saving it as a Skill, you keep your Project instructions clean and invoke the structured format only when you need it. Skills also work across all your projects, so you can reuse this in other clinical workspaces.

Test with a Sample Case

Now test your new Skill with a textbook clinical vignette.

  • Go back to your projects.
  • Start a new chat in your Clinical Assistant project.
  • Click on the + (plus icon) to activate your Differential Diagnosis skill from the skill picker.
  • Paste in a clinical case like this:
45-year-old male presents to the ED with acute onset chest pain, started
30 minutes ago. Pain is substernal, radiating to the left arm. Patient is
diaphoretic and nauseous. History of hypertension and smoking (20 pack-years).
Vitals: BP 160/95, HR 110, SpO2 96%.
  • Press Enter and review Claude's response.

Check whether the format matches the structure you defined, whether the likelihood levels are reasonable for the presented case, and whether the red flags are clinically appropriate. The "Cannot-Miss Diagnoses" section should include dangerous conditions that must be ruled out.

What if the format breaks?

Structured output formatting may not hold perfectly every time. If Claude deviates from your defined format, re-send your message or start a new chat with the Skill activated again.

Iterate and Refine

A straightforward case is one thing. Real clinical reasoning shines when dealing with ambiguity. Try a more challenging case to test how Claude handles uncertainty.

  • Start a new chat in your Clinical Assistant project with the Differential Diagnosis skill activated.
  • Paste a more ambiguous case:
30-year-old female presents with 3 months of progressive fatigue, unintentional
weight loss of 15 lbs, and intermittent low-grade fevers. Labs show mild anemia
and elevated ESR. No significant past medical history. Physical exam is
unremarkable except for mild pallor.
  • Press Enter and review the output.

Notice how Claude handles the ambiguity. With a vague presentation like this, there are many possible differentials and the likelihood levels should reflect genuine uncertainty.

Now let's refine the skill so Claude asks clarifying questions before jumping to conclusions when information is limited.

  • In your current conversation, send Claude a message asking it to update your skill with these new rules:
Update my Differential Diagnosis skill to add these rules:

Additional rules for differential diagnosis:
- If the presentation is insufficient to rank diagnoses, ask me clarifying
  questions before generating the differential.
- Always include at least one "cannot miss" diagnosis even if unlikely.
  • Claude will update the skill for you. Confirm the changes when prompted.

What do these refinements do?

The first rule teaches Claude to behave like a real clinician: gather more information before forming conclusions. Instead of guessing with incomplete data, it will ask targeted follow-up questions. The second rule ensures dangerous conditions like pulmonary embolism or malignancy are always flagged, even if they seem unlikely based on the initial presentation.

  • Start a new chat with the skill activated and test the ambiguous case again.

I got this! 💪

Nice work! Once you have your new chat open with the skill activated, paste the ambiguous case and press Enter:

30-year-old female presents with 3 months of progressive fatigue, unintentional
weight loss of 15 lbs, and intermittent low-grade fevers. Labs show mild anemia
and elevated ESR. No significant past medical history. Physical exam is
unremarkable except for mild pallor.

Walk me through it

No worries, here is how to start a new chat with the skill activated:

  • Go back to your projects.
  • Start a new chat in your Clinical Assistant project.
  • Click on the + (plus icon) to activate your Differential Diagnosis skill from the skill picker.
  • Paste the ambiguous case and press Enter:
30-year-old female presents with 3 months of progressive fatigue, unintentional
weight loss of 15 lbs, and intermittent low-grade fevers. Labs show mild anemia
and elevated ESR. No significant past medical history. Physical exam is
unremarkable except for mild pallor.

Claude should now ask clarifying questions before jumping to a differential when the information is insufficient. This is a much safer clinical behavior.

Why add the clarifying questions rule?

In real clinical practice, a physician would ask follow-up questions before forming a differential. Teaching Claude to do the same prevents premature conclusions and models better clinical reasoning.

Your differential diagnosis Skill is ready. Next, take on the Secret Mission to add a clinical safety layer that makes your assistant cite sources, flag uncertainty, and refuse to diagnose.

Secret mission

Your Clinical Assistant works well, but would you trust it in a real clinical workflow without safety guardrails? In this secret mission, you will add a safety layer that makes your assistant cite sources, flag uncertainty, and refuse to make definitive diagnoses.

This combines chain-of-thought verification with the role prompting, audience adaptation, and structured output techniques from earlier steps. The result is the most professionally valuable outcome of the entire project: a clinical AI assistant that knows its own limits.

In this secret mission, get ready to:

  • Add citation requirements to the Project instructions.
  • Add explicit uncertainty flagging rules.
  • Add a "refusal to diagnose" guardrail.
  • Test the complete safety layer with a clinical scenario.

Build a Clinical Safety Layer

Clean Up Your Resources

Clean Up Your Resources

Decide whether to keep your resources running, pause them to come back later, or delete them entirely. This project runs entirely in Claude.ai with no ongoing costs.

Resources you used:

  • Claude.ai free account.
  • "Clinical Assistant" Project with custom instructions.
  • Conversation history within the project.

✔️ Keep everything running

No action needed. Choose this if you're still actively building or want to keep testing right away.

Your Claude Project and all conversation history persist automatically at no cost. This is the recommended option since the Clinical Assistant is designed to be a reusable tool.

✋ Pause - I'll come back to this later

No action needed. Claude Projects persist automatically, so you can return anytime.

Your custom instructions and conversation history will be waiting for you when you come back.

ⓧ Delete - I don't want to use this again

Remove your project if you want a clean slate.

  • Open your "Clinical Assistant" Project in Claude.ai.
  • Click the settings or gear icon.
  • Select Delete Project.

This removes all custom instructions and conversation history permanently.

Well done!

Well done!

Nice work! 🚀 You have built a fully configured Clinical AI Assistant in Claude with safety guardrails, optimized prompts, and structured outputs.

You've learned how to:

  • ⚒️ Set up a Claude Project with persistent clinical context and role instructions.
  • 💊 Use role prompting to check drug interactions like a clinical pharmacist.
  • 📖 Create Claude Styles with audience adaptation to rewrite medical content at multiple reading levels.
  • 🩺 Build a Claude Skill with structured output for differential diagnosis with confidence levels and red flags.
  • 💎 Add a clinical safety layer with citations, uncertainty flagging, and diagnosis refusal.

Ready to quiz yourself?

p.s. Does it say "Still tasks to complete!" at the bottom of the screen?

This means you still have screenshots left to upload, or questions left to answer!

  1. Press Ctrl+F (Windows) or Command+F (Mac) on your keyboard.
  2. Search for the text Return to later.
  3. Jump straight to your incomplete tasks!
  4. 🙋‍♀️ Still stuck? Ask the community!