Clinical Note Writing with AI
Learn to effectively use AI for drafting H&Ps, progress notes, treatment summaries, and other clinical documentation.
- Structure effective prompts for clinical documentation
- Generate draft clinical notes efficiently
- Establish a verification workflow
- Adapt AI outputs to institutional templates
Clinical documentation is one of the most time-consuming aspects of medical practice. AI can help reduce documentation burden while maintaining quality—when used correctly.
The Role of AI in Documentation:
AI is best used as a drafting assistant, not an autonomous documentation system. The workflow is:
- You provide context (scenario, template, requirements)
- AI generates a draft
- You review, verify, and edit
- You finalize with patient-specific accuracy
Benefits:
- Faster first drafts
- Consistent structure and formatting
- Reduced cognitive load on routine documentation
- More time for complex clinical reasoning
Risks to Mitigate:
- Over-reliance leading to errors
- Propagating AI-generated inaccuracies
- Privacy violations if PHI is included in prompts
The quality of AI output depends heavily on prompt quality. Follow this structure for clinical documentation:
The CRAFT Framework:
C - Context: Set the scene
- "You are helping a radiation oncologist draft clinical documentation."
- "This is for a [type of note] in [setting]."
R - Role & Requirements: Specify what you need
- "Generate a treatment summary following this template..."
- "The note should include: [specific sections]"
A - Audience: Who will read this?
- "This will be sent to the referring physician."
- "This is for the patient's permanent medical record."
F - Format: Structure and style
- "Use bullet points for the assessment."
- "Keep it concise, under 500 words."
- "Follow standard SOAP format."
T - Task: The specific request
- "Draft a note for a patient who completed..."
- "Summarize the following treatment course..."
Example Prompt:
Context: You are assisting a radiation oncologist with documentation.
Task: Draft a brief end-of-treatment summary for documentation purposes.
Scenario: A 62-year-old male completed definitive chemoradiation for Stage IIIA NSCLC. Treatment: 60 Gy in 30 fractions via IMRT with concurrent weekly carboplatin/paclitaxel. Treatment was completed without breaks. He experienced Grade 2 esophagitis managed with viscous lidocaine and Grade 1 fatigue.
Format: Use standard summary format with sections for Diagnosis, Treatment Delivered, Tolerance, and Follow-up Plan. Keep it concise.Prompt Used:
"Draft an end-of-treatment summary for: 58-year-old female who completed adjuvant whole breast radiation following lumpectomy for T1N0 ER+/PR+/HER2- invasive ductal carcinoma. Treatment: 40 Gy in 15 fractions to whole breast via IMRT, followed by 10 Gy boost in 4 fractions to lumpectomy cavity. Completed without treatment breaks. Mild Grade 1 skin erythema in inframammary fold at completion. She will continue on adjuvant anastrozole."
AI-Generated Draft:
RADIATION ONCOLOGY END-OF-TREATMENT SUMMARY
Patient: [Name]
Date of Completion: [Date]
Diagnosis:
T1N0 ER+/PR+/HER2- invasive ductal carcinoma of the left breast, status post lumpectomy
Treatment Delivered:
The patient completed adjuvant whole breast radiation therapy as follows:
- Whole breast: 40 Gy in 15 fractions via IMRT
- Lumpectomy cavity boost: 10 Gy in 4 fractions
- Total dose to tumor bed: 50 Gy
- Treatment completed without unplanned breaks
Treatment Tolerance:
Treatment was well tolerated. At completion, the patient exhibited Grade 1 skin erythema localized to the inframammary fold, expected to resolve within 2-4 weeks with supportive care.
Current Medications:
Anastrozole 1 mg daily (adjuvant endocrine therapy, ongoing)
Follow-up Plan:
- Radiation oncology follow-up in 4-6 weeks to assess skin healing
- Continue care with medical oncology for endocrine therapy monitoring
- Annual mammography per guidelines
Review Notes:
The physician would then:
- Add actual patient name and dates
- Verify the staging matches records
- Confirm dose/fractionation from treatment records
- Add any patient-specific details or concerns
- Adjust follow-up plan to institutional practice
Every AI-generated note requires verification before use.
Verification Checklist:
Common AI Errors in Clinical Notes:
- Inventing specific dates or timeframes
- Adding plausible-sounding but incorrect details
- Using different toxicity grading than intended
- Including standard recommendations that don't apply to this patient
- Formatting that doesn't match your EMR template
Time-Saving Tip:
Create verified base templates for common scenarios. Use AI to adapt these templates rather than generating from scratch each time. This reduces the verification burden.
Original Prompt:
"Write a note for a prostate cancer patient who finished radiation."
This prompt is too vague. Let's improve it using the CRAFT framework.
Improved Prompt:
Context: You are helping a radiation oncologist draft an end-of-treatment summary for the medical record.
Task: Create a treatment completion summary for a prostate cancer patient.
Scenario:
- 68-year-old male
- Diagnosis: Intermediate-risk prostate adenocarcinoma (Gleason 3+4=7, PSA 12, T2b)
- Treatment: Definitive IMRT, 78 Gy in 39 fractions to prostate and proximal seminal vesicles
- Concurrent ADT: Lupron, started 2 months prior, planned for total 6 months
- Tolerance: Grade 1 urinary frequency, no GI toxicity
- Completed without breaks
Format:
- Standard summary with sections: Diagnosis, Treatment, Tolerance, Recommendations
- Concise, suitable for EMR documentation
- Professional medical terminologyWhy This Works Better:
- Provides specific clinical details
- Specifies the exact treatment parameters
- Includes toxicity information with grading
- Requests specific format and sections
- Sets appropriate tone (professional, for medical record)
Try It: Practice rewriting vague prompts into structured ones using the CRAFT framework.
Key Takeaways:
- Use the CRAFT framework for structured, effective prompts: Context, Role/Requirements, Audience, Format, Task.
- AI generates drafts, not final notes. Always verify against source records.
- Create reusable templates for common note types to reduce verification burden.
- Establish a verification checklist and use it consistently.
- Never include PHI in prompts. Use fictional scenarios or de-identified placeholders.
Practical Application:
Start with low-stakes documents (patient education materials, template structures) before using AI for clinical documentation. Build your verification habits on simple tasks first.
Next Module: Patient Communication - Learn to create patient-friendly materials explaining radiation therapy.