For Medical Scribes ·
What you'll accomplish
By the end of this guide, you'll have a systematic workflow for using Claude to review and edit clinical notes generated by ambient AI tools (like Nuance DAX, Abridge, or DeepScribe) — catching AI errors, hallucinations, and missing elements before the physician sees the note. This positions you as a quality professional, not just a typist — and it's the most future-proof skill a medical scribe can develop in 2025.
What you'll need
Open claude.ai and start a new conversation. Begin with this context-setting message:
"I'm a medical scribe reviewing AI-generated clinical notes in [specialty] before physician attestation. My job is to catch: missing documentation elements, clinical inconsistencies, overly generic or AI-sounding language, and errors in the Assessment & Plan. Help me review notes as a quality checker — flag issues clearly so the physician can review them."
What you should see: Claude confirms it's ready to act as a clinical note quality reviewer.
The most powerful thing you can do is establish a consistent review framework. Ask Claude:
"Create a clinical note quality checklist for [specialty] encounter documentation. Include checks for: HPI completeness (OLDCART elements), ROS completeness by relevant system, Physical Exam documentation accuracy, A&P clinical logic (does the plan match the assessment?), medication documentation accuracy, and red flags for AI hallucination."
What you should see: A structured checklist with specific items for each note section. Save this checklist — you'll use it for every note review. Troubleshooting: If the checklist is too generic, add: "Make this specific to [specialty] — what are the documentation requirements unique to this specialty?"
Find an AI-generated note from your system (or create a realistic test note). Copy it completely. In Claude, type:
"Here is an AI-generated [specialty] encounter note. Review it against these criteria: [paste your checklist]. Flag every issue you find with: [ISSUE TYPE] — what the note says vs. what it should say."
Paste the full note below.
What you should see: Claude returns a structured review with flagged issues like:
Ask Claude: "What are the most common ways AI ambient documentation tools make errors or hallucinate in [specialty] notes? Give me 10 specific red flags to check in every note."
What you should see: A specialty-specific list of common AI errors. For cardiology, this might include: confusing medication names (e.g., carvedilol vs. bisoprolol), hallucinating normal lab values, or documenting exam findings that the patient history makes clinically inconsistent.
For each common issue type, have Claude give you a correction template. For example: "Give me template correction language for when an AI note's HPI is missing duration of symptoms. What should I type in the note margin or flag to the physician?"
What you should see: Standard flag language like: "[PHYSICIAN REVIEW NEEDED — HPI incomplete: Duration of [chief complaint] not documented. Please verify and add duration before attestation.]"
Create a consistent process: (1) AI generates note after encounter, (2) You copy the note into Claude for a 2-3 minute review, (3) You return to the EHR and make corrections or add physician review flags, (4) Physician reviews the improved note and attests.
Initial note review:
Review this [specialty] AI-generated encounter note. Flag issues with: [ISSUE TYPE] — what it says vs. what it should say. Focus on completeness, clinical logic, and hallucinations.
[paste note]
Completeness check:
Check this [specialty] note for missing required documentation elements. Use [specialty] documentation standards. List every missing element with the section it belongs in.
[paste note]
Clinical logic check:
Does the Assessment match the Plan in this note? Flag any disconnect between the documented assessment and the treatment/orders in the Plan.
[paste note]
Hallucination detection:
Review this AI-generated note for potential hallucinations — clinical findings, lab values, or historical elements that may have been invented by the AI rather than documented in the actual encounter. Flag anything that seems inconsistently specific or uncorroborated.
[paste note]