jixiaxue 知识库
evidence · 2026-04-15

T0-3-real-world-02-medical-prompt

/Users/shanfang/Documents/pe/jixiaxuegong/research/提示工程教程/evidence/T0-anthropic/T0-3-real-world-02-medical-prompt.md

来源：https://github.com/anthropics/courses/blob/master/real_world_prompting/02_medical_prompt.ipynb 爬取日期：2026-03-22

Lesson 2: A real-world prompt

In the previous lesson, we discussed several key prompting tips and saw an example of how to use each in isolation. Let’s now try writing a much larger prompt that incorporates many of the techniques we just covered.

Our prompting goal

This lesson will focus on writing a medical record summarizer prompt that takes in long medical records and generates a summary containing important information to assist doctors in preparing for upcoming appointments.

Each patient medical record looks something like this:

Patient Name: Evelyn Thompson
Age: 78
Medical Record:

1985: Diagnosed with type 2 diabetes, started on metformin
1992: Developed hypertension, prescribed lisinopril
1998: Total hip replacement (right) due to osteoarthritis
2000: Diagnosed with hypothyroidism, started on levothyroxine
2003: Cataract surgery (both eyes)
2005: Admitted for atrial fibrillation, started on warfarin
2008: Vitamin B12 deficiency diagnosed, monthly injections started
2010: Increased metformin dose due to rising A1C levels
2011: Admitted for transient ischemic attack (TIA), added aspirin to regimen
2013: Diagnosed with stage 2 breast cancer, underwent lumpectomy and radiation
2014: Started on anastrozole for breast cancer recurrence prevention
2015: Developed chronic kidney disease (CKD) stage 3, metformin adjusted
2017: Total knee replacement (left) due to osteoarthritis
2018: Hospitalized for pneumonia, treated with IV antibiotics
2019: Mild cognitive impairment noted, started on donepezil
2020: Lisinopril dosage increased due to refractory hypertension
2021: Recurrent UTIs, prescribed low-dose prophylactic antibiotics
2022: Annual mammogram clear, but eGFR shows worsening kidney function
2023: Mobility declining, started physical therapy and home health aide visits

Our end goal is to generate consistent record summaries to help providers prepare for upcoming appointments. Each summary should contain key pieces of information including:

An example output for the above medical record might look something like this:

Name: Evelyn Thompson
Age: 78

Key Diagnoses:
- Type 2 Diabetes (1985)
- Hypertension (1992)
- Osteoarthritis (Hip and Knee Replacements in 1998 and 2017)
- Hypothyroidism (2000)
- Atrial Fibrillation (2005)
- Vitamin B12 Deficiency (2008)
- Transient Ischemic Attack (TIA) (2011)
- Breast Cancer (2013)
- Chronic Kidney Disease (CKD) Stage 3 (2015)
- Pneumonia (2018)
- Mild Cognitive Impairment (2019)
- Recurrent Urinary Tract Infections (UTIs) (2021)

Medications:
- Metformin (Diabetes)
- Lisinopril (Hypertension)
- Levothyroxine (Hypothyroidism)
- Warfarin (Atrial Fibrillation)
- Aspirin (Antiplatelet)
- Anastrozole (Breast Cancer Recurrence Prevention)
- Donepezil (Cognitive Impairment)
- Low-dose Prophylactic Antibiotics (Recurrent UTIs)

Other Treatments:
- Total Hip Replacement (1998)
- Cataract Surgery (2003)
- Vitamin B12 Injections (2008)
- Lumpectomy and Radiation (Breast Cancer, 2013)
- Total Knee Replacement (2017)
- Physical Therapy and Home Health Aide (2023)

Recent Concerns:
- Worsening Kidney Function (eGFR Decline in 2022)
- Declining Mobility (2023)

Action Items:
- Monitor Kidney Function and Adjust Medications as Needed
- Continue Physical Therapy and Home Health Support
- Evaluate for Cognitive Decline and Adjust Treatment Plan
- Address Mobility Issues and Fall Risk
- Ensure Adherence to Recommended Cancer Screening

Test data: Patient Records

Here’s a Python list containing 5 medical records that we’ll try our prompt with:

patient_records = [
    """
Patient Name: Evelyn Thompson
Age: 78
Medical Record:

1985: Diagnosed with type 2 diabetes, started on metformin
1992: Developed hypertension, prescribed lisinopril
1998: Total hip replacement (right) due to osteoarthritis
2000: Diagnosed with hypothyroidism, started on levothyroxine
2003: Cataract surgery (both eyes)
2005: Admitted for atrial fibrillation, started on warfarin
2008: Vitamin B12 deficiency diagnosed, monthly injections started
2010: Increased metformin dose due to rising A1C levels
2011: Admitted for transient ischemic attack (TIA), added aspirin to regimen
2013: Diagnosed with stage 2 breast cancer, underwent lumpectomy and radiation
2014: Started on anastrozole for breast cancer recurrence prevention
2015: Developed chronic kidney disease (CKD) stage 3, metformin adjusted
2017: Total knee replacement (left) due to osteoarthritis
2018: Hospitalized for pneumonia, treated with IV antibiotics
2019: Mild cognitive impairment noted, started on donepezil
2020: Lisinopril dosage increased due to refractory hypertension
2021: Recurrent UTIs, prescribed low-dose prophylactic antibiotics
2022: Annual mammogram clear, but eGFR shows worsening kidney function
2023: Mobility declining, started physical therapy and home health aide visits
    """,
    """
Patient Name: Marcus Reyes
Age: 42
Medical Record:

2001: Diagnosed with generalized anxiety disorder (GAD), started on paroxetine
2003: Diagnosed with major depressive disorder (MDD), added bupropion
2005: Hospitalized for suicidal ideation, added cognitive behavioral therapy (CBT)
2007: Diagnosed with attention-deficit/hyperactivity disorder (ADHD), started on methylphenidate
2009: Reported side effects from paroxetine, switched to escitalopram
2012: Diagnosed with obstructive sleep apnea (OSA), started CPAP therapy
2014: Diagnosed with hypertension, started on losartan
2015: Weight gain noted, referred to nutritionist
2016: Diagnosed with type 2 diabetes, started on metformin
2017: Hospitalized for diabetic ketoacidosis (DKA), insulin therapy initiated
2018: Reported nightmares, switched from bupropion to venlafaxine
2019: Gastroesophageal reflux disease (GERD) diagnosis, started on omeprazole
2020: Divorce, increased therapy sessions, added dialectical behavior therapy (DBT)
2021: Developed plantar fasciitis, prescribed orthotics and physical therapy
2022: Admitted for panic attack, mistaken for myocardial infarction, cardiac workup negative
2023: Attempted suicide, inpatient psychiatric treatment for 30 days
2023: Post-discharge, started on new antipsychotic (quetiapine) and mood stabilizer (lamotrigine)
2024: Reports improvement in mood and sleep, weight loss noted
2024: A1C levels improved, insulin dose decreased
    """,
    """
Patient Name: Lily Chen
Age: 8
Medical Record:

2016 (Birth): Born at 34 weeks, diagnosed with Tetralogy of Fallot (TOF)
 - Immediate surgery to place a shunt for increased pulmonary blood flow
2016 (3 months): Echocardiogram showed worsening right ventricular hypertrophy
2017 (8 months): Complete repair of TOF (VSD closure, pulmonary valve replacement, RV outflow tract repair)
2017 (10 months): Developed post-operative arrhythmias, started on amiodarone
2018 (14 months): Developmental delay noted, referred to early intervention services
2018 (18 months): Speech therapy initiated for delayed language development
2019 (2 years): Diagnosed with failure to thrive, started on high-calorie diet
2019 (2.5 years): Occupational therapy started for fine motor skill delays
2020 (3 years): Cardiac catheterization showed mild pulmonary stenosis
2020 (3.5 years): Diagnosed with sensory processing disorder (SPD)
2021 (4 years): Started integrated preschool program with IEP (Individualized Education Plan)
2021 (4.5 years): Hospitalized for RSV bronchiolitis, required brief oxygen support
2022 (5 years): Echocardiogram showed progression of pulmonary stenosis, balloon valvuloplasty performed
2022 (5.5 years): Diagnosed with attention-deficit/hyperactivity disorder (ADHD), started behavioral therapy
2023 (6 years): Cochlear implant surgery for sensorineural hearing loss
2023 (7 years): Started mainstream school with continued IEP support
2024 (7.5 years): Occupational therapy discontinued, met fine motor skill goals
2024 (8 years): Periodic cardiac follow-up shows stable pulmonary valve function
2024 (8 years): Speech development progressing well, ongoing therapy
    """,
    """
Patient Name: Jason Tran
Age: 25
Medical Record:

2010 (11 yrs): Diagnosed with asthma, started on albuterol inhaler
2012 (13 yrs): First football concussion, brief loss of consciousness
2013 (14 yrs): Fractured right tibia during soccer, surgical fixation
2014 (15 yrs): Second concussion, resulting in post-concussion syndrome
 - Symptoms: headaches, dizziness, memory problems
 - Referred to pediatric neurologist, cognitive rehabilitation therapy
2015 (16 yrs): Developed anxiety and depression, started on fluoxetine
2016 (17 yrs): ACL tear (left knee) during basketball, reconstructive surgery
 - 6-month rehabilitation, switched to non-contact sports
2017 (18 yrs): Graduated high school, started college on academic scholarship
2018 (19 yrs): Diagnosed with PTSD related to sports injuries
 - Started cognitive-behavioral therapy (CBT)
2019 (20 yrs): Tried to return to basketball, experienced panic attack
 - Increased therapy sessions, added exposure therapy
2020 (21 yrs): COVID-19 pandemic, remote learning, reported increased anxiety
 - Started mindfulness meditation and yoga
2021 (22 yrs): Diagnosed with sleep apnea, started CPAP therapy
 - Sleep study suggested link between concussions and sleep disorder
2022 (23 yrs): Gradual return to low-impact sports (swimming, cycling)
 - Reported improved mood and sleep quality
2023 (24 yrs): Graduated college, started job in sports analytics
 - Continuing therapy, now biweekly
 - Volunteering with youth concussion awareness program
2024 (25 yrs): Annual check-up - asthma well-controlled, mental health stable
 - No sports-related injuries in past 2 years
 - Training for first half-marathon
    """,
    """
Patient Name: Amira Khan
Age: 36
Medical Record:

2011: Recurrent joint pain and fatigue, initial diagnosis of fibromyalgia
 - Started on pregabalin and physical therapy
2012: Developed persistent rash and photosensitivity
 - Dermatologist diagnosed cutaneous lupus (CLE)
 - Started on topical corticosteroids and sunscreen
2013: Complained of severe fatigue, hair loss, and cognitive issues ("brain fog")
 - Blood tests showed positive ANA and anti-dsDNA antibodies
 - Diagnosed with Systemic Lupus Erythematosus (SLE)
 - Started on hydroxychloroquine and low-dose prednisone
2014: Hospitalized for lupus nephritis (class III)
 - Renal biopsy confirmed diagnosis
 - Started on mycophenolate mofetil and increased prednisone
2015: Developed interstitial lung disease (ILD) secondary to SLE
 - Started on cyclophosphamide pulse therapy
 - Required home oxygen therapy
2016: Diagnosed with secondary Sjogren's syndrome
 - Symptoms: dry eyes, dry mouth
 - Started on pilocarpine and artificial tears
2017: Hospitalized for lupus cerebritis
 - Symptoms: seizures, confusion, memory loss
 - MRI showed brain inflammation
 - Treated with high-dose steroids and rituximab
2018: Developed avascular necrosis (AVN) of hip, steroid-induced
 - Underwent total hip replacement surgery
2019: Started on belimumab to reduce flare frequency
 - Gradual reduction in prednisone dosage
2020: COVID-19 pandemic, self-isolated due to immunosuppression
 - Telemedicine follow-ups, home infusions
2021: Diagnosed with steroid-induced diabetes
 - Started on metformin and insulin
 - Referred to endocrinologist and nutritionist
2022: Flare-up of lupus, increased joint pain and fatigue
 - Adjusted medications: increased mycophenolate, added abatacept
2023: Cardiovascular screening due to long-term steroid use
 - Echo showed early signs of diastolic dysfunction
 - Started on ACE inhibitors, referred to cardio-rehab
2024: Improvement noted in all organ systems
 - Tapering immunosuppressants, monitoring closely
 - Continues physical therapy, yoga for joint health
 - Planning pregnancy, consulted with high-risk OB
    """
]

Our initial “bad” prompt

Let’s start with a very simple prompt that asks Claude to generate us a summary. A very simple first attempt might look something like this:

I have this patient medical record. Can you summarize it for me?

{medical record goes here}

I need this for a quick review before the patient’s appointment tomorrow.

initial_prompt = """
I have this patient medical record. Can you summarize it for me?

{record}

I need this for a quick review before the patient's appointment tomorrow.
"""
from anthropic import Anthropic
from dotenv import load_dotenv

load_dotenv()
client = Anthropic()

def generate_summary_with_bad_prompt(patient_record):
    prompt_with_record = initial_prompt.format(record=patient_record)
    response = client.messages.create(
        model="claude-3-sonnet-20240229",
        max_tokens=4096,
        messages=[{"role": "user", "content": prompt_with_record}]
    )
    print("===============================")
    print(response.content[0].text)

Testing the bad prompt

Results from running all 5 records showed that the summaries are all over the place:

Improving the prompt

As we learned in the previous lesson, there are a clear set of prompting techniques that could help us get better and more consistent results, including:

In upcoming lessons, we’ll discuss a specific, nuanced approach to prompt engineering and selecting prompting techniques. In this lesson we’ll take a “shotgun” approach and use all of them at once.

Adding a system prompt

system = """
You are a highly experienced medical professional with a specialty in translating complex patient histories into concise, actionable summaries.
Your role is to analyze patient records, identify critical information, and present it in a clear, structured format that aids in diagnosis and treatment planning.
Your summaries are invaluable for busy healthcare providers who need quick insights into a patient's medical history before appointments.
"""

Structuring input data

One of the most important prompting tips when working with Claude is to clearly label your input data using XML tags.

updated_prompt = """
<patient_record>
{record}
</patient_record>
"""

Provide clear instructions

updated_prompt = """
I need your help summarizing patient medical records for our team of doctors.
We have a series of follow-up appointments tomorrow, and the doctors need quick, insightful summaries to prepare.

Each summary should include the following elements in this order:
- The patient's name
- The patients age
- A bulleted list of key diagnoses in chronological order
- A bulleted list of medications the patient is prescribed
- A bulleted list of other treatments: non-medication treatments like CBT or physical therapy
- A short bulleted list of recent concerns
- A bulleted list of key action items to help our doctors prepare for the upcoming patient visit

<patient_record>
{record}
</patient_record>
"""

Adding examples

Examples are one of the most powerful tools for enhancing Claude’s performance. By providing a few well-crafted examples, we can significantly improve accuracy, consistency, and quality.

Example input (Ethan Blackwood, 55-year-old) and corresponding output were added to the prompt inside <example> tags.

Output XML structure

Adding <summary> tags to the example output teaches Claude to wrap its output in XML tags, making it easy to parse.

Final improved prompt

system = """
You are a highly experienced medical professional with a specialty in translating complex patient histories into concise, actionable summaries.
Your role is to analyze patient records, identify critical information, and present it in a clear, structured format that aids in diagnosis and treatment planning.
Your summaries are invaluable for busy healthcare providers who need quick insights into a patient's medical history before appointments.
"""

updated_prompt = """
I need your help summarizing patient medical records for our team of doctors.
We have a series of follow-up appointments tomorrow, and the doctors need quick, insightful summaries to prepare.

Each summary should include the following elements in this order:
- The patient's name
- The patients age
- A bulleted list of key diagnoses in chronological order
- A bulleted list of medications the patient is prescribed
- A bulleted list of other treatments: non-medication treatments like CBT or physical therapy
- A short bulleted list of recent concerns
- A bulleted list of key action items to help our doctors prepare for the upcoming patient visit

Here's an example of how we'd like the summaries formatted:

<example>
<patient_record>
Patient Name: Ethan Blackwood
Age: 55
Medical Record:

2010: Annual check-up, mild hypertension noted
 - Started on lifestyle modifications (diet, exercise)
2012: Diagnosed with moderate depression following job loss
 - Started on sertraline and cognitive-behavioral therapy (CBT)
2014: New job, reported improved mood
 - Continued sertraline, reduced CBT sessions
2015: Mild back pain, diagnosed with early degenerative disc disease
 - Physical therapy and over-the-counter NSAIDs prescribed
2016: Hypertension worsened, started on lisinopril
2017: Routine colonoscopy showed benign polyps, removed during procedure
2018: Developed persistent cough, chest X-ray clear
 - Diagnosed with Gastroesophageal Reflux Disease (GERD)
 - Started on omeprazole
2019: Diagnosed with obstructive sleep apnea (OSA)
 - Started CPAP therapy, reported improved energy levels
2020: COVID-19 pandemic, worked from home
 - Reported increased anxiety, CBT sessions resumed (telehealth)
 - COVID-19 vaccination (Moderna, both doses)
2021: Mild knee pain, MRI showed minor meniscus tear
 - Arthroscopic surgery recommended, patient opted for conservative management
2022: Annual check-up showed pre-diabetes (A1C: 6.1%)
 - Intensified lifestyle modifications, referred to nutritionist
- Discontinued omeprazole due to resolved GERD symptoms
2023: Blood tests showed elevated PSA (Prostate-Specific Antigen)
 - Prostate biopsy performed, results negative for cancer
- Knee pain worsened, agreed to arthroscopic surgery
2024: Post-op knee recovery: good, continuing physical therapy
 - A1C levels improved (5.8%), pre-diabetes resolved
 - Stress test normal, but mild LVH on echocardiogram
 - Started on low-dose ACE inhibitor for cardioprotection
</patient_record>

<summary>
Name: Ethan Blackwood
Age: 55

Key Diagnoses:
- Hypertension (2010)
- Depression (2012)
- Degenerative Disc Disease (2015)
- Gastroesophageal Reflux Disease (GERD) (2018)
- Obstructive Sleep Apnea (OSA) (2019)
- Pre-diabetes (2022)
- Meniscus Tear (2021)
- Left Ventricular Hypertrophy (LVH) (2024)

Medications:
- Sertraline (depression)
- Lisinopril (hypertension)
- Omeprazole (GERD) - discontinued in 2022
- Low-dose ACE inhibitor (cardioprotection - 2024)

Other Treatments:
- Cognitive Behavioral Therapy (CBT) (depression)
- Physical therapy (back pain, post-op knee recovery)
- CPAP therapy (OSA)
- Arthroscopic knee surgery (2023)

Recent Concerns:
- Worsening knee pain
- Elevated PSA (2023)
- Left ventricular hypertrophy on echocardiogram (2024)

Action Items:
- Follow up on post-op knee recovery and physical therapy
- Monitor PSA levels and prostate health
- Optimize blood pressure and hypertension management
- Assess need for further cardiac workup after LVH finding
</summary>
</example>

Now, please summarize the following patient record in the same format:

<patient_record>
{record}
</patient_record>
"""

Recap of the prompt changes

Testing results

With the improved prompt, all 5 patient record summaries now follow the exact same structure and format consistently. Each output includes:

Switching things up: JSON!

Often we want specific structured responses that are easier to programmatically digest. The most common approach is to use JSON.

Note: The easiest way to ‘force’ a JSON response is through Claude’s tool use functionality, which is covered in a separate lesson.

To adapt the prompt for JSON output:

updated_json_prompt = """
I need your help summarizing patient medical records for our team of doctors.
We have a series of follow-up appointments tomorrow, and the doctors need quick, insightful summaries to prepare.

Please provide these summaries in JSON format with the following structure:
{
    "name": "Patient's full name",
    "age": patient's age as an integer,
    "key_diagnoses": [
        {
            "diagnosis": "Primary diagnosis",
            "year": year of diagnosis as an integer
        },
        ...
    ],
    "medications": [
        {
            "name": "Medication name",
            "purpose": "Brief description of what it's for"
        },
        ...
    ],
    "other_treatments": [
        {
            "treatment": "Treatment name",
            "purpose": "Brief description of what it's for"
        },
        ...
    ],
    "recent_concerns": [
        "Brief statement of recent health issue or concern"
    ],
    "action_items": [
        "Action item 1",
        "Action item 2",
        ...
    ]
}
"""

Summary

This notebook demonstrates a comprehensive approach to prompt engineering for medical record summarization, showing:

  1. Initial Poor Prompt - A vague, unstructured prompt that produces inconsistent outputs
  2. Incremental Improvements - Adding system prompts, XML tags, clear instructions, examples, and output formatting
  3. Testing and Results - Comparing outputs from the poor prompt vs. improved prompt across 5 different patient medical records
  4. Advanced Formatting - Transitioning from text output to JSON for programmatic access

The lesson emphasizes that by applying multiple prompting techniques together, Claude produces much more consistent, structured, and reliable outputs suitable for real-world applications.