Residents keep telling me the same thing: they tried ChatGPT for notes, then spent longer editing than if they'd just written the SOAP themselves.

The problem isn't the model. It's the prompt.

"Write a SOAP note for chest pain" produces a generic, perfectly structured, slightly wrong note. Attendings can smell it. So can patients if you reuse the education paragraph.

Three rules that actually work:

1. Force the format your team already uses. "Use this exact heading order: CC, HPI (OLDCARTS), ROS pertinent only, A/P as problem list with one-line rationale. Do not invent negatives I didn't give you."

1. Dump the messy facts, not the diagnosis. Age, timing, what you actually heard on exam, what you already ordered. Let the model organize. You still decide.

1. Ban the voice. "No 'it is important to note,' no 'comprehensive evaluation,' no 'further management as clinically indicated.' Write like a tired resident at 6:40pm."

That's the whole ebook: 500 of those prompts, sorted by what you actually type after clinic — SOAP, H&P, dispo, patient education, handoffs — so you're not rebuilding the prompt every night.