03 · What leaves the building · 35 min
The AIs
The first question is not which model is best. It is whether this text is allowed to leave the machine in front of you.
Three zones
- 01 Hospital machine, no cloud
Anything that still has a name, MRN, face, voice, full imaging, or a rare story plus a date. Dictation counts. A plugin that “only summarizes” counts.
- 02 De-identified cloud
Drafts, outlines, code for a teaching site, video scripts, after the case has been rewritten clean. Claude, ChatGPT, Gemini, Grok — pick by the job, then actually check the output.
- 03 Local model
When the file cannot leave the laptop and you still want a draft. Usually via something like Ollama. Weaker writing. Correct trade when privacy is the point. Not a first-weekend project.
Which door is this
Strip identifiers on a machine allowed to hold the chart, or don't use a model on that note. A cloud chat is the wrong place to discover you forgot the sticker.
Ask for a cut to a time limit, not a rewrite into something smoother. Keep the decision. Delete the throat-clearing.
A long-context chat is a sorting desk, not a bibliography. If a citation matters, you retrieve it. A model reference list is a lead.
Good for a temperature check. Bad as a source. If you repeat a claim, you should be able to point at the paper or the data yourself.
You name the structure and the move. The model does not get to invent anatomy it cannot see. Record the words after, in a quiet room.
A local model is a privacy tool. It is slower and usually worse at long writing. That trade is correct when the alternative is sending the file out.
Never generate a picture of a real patient's anatomy and call it the case. A schematic is a schematic. Say so on the page.
What to open for which job
Pay for one writing model you will actually open when you are tired. A second model is for when the first one is fluently wrong. Four subscriptions you rotate is not a system.
| Job | Use | Do not |
|---|---|---|
| A long de-identified talk or a project you return to | One writing model with a saved project and the brief below | A fresh chat that has forgotten your rules |
| A pile of PDFs | Whichever app lets you attach them — then open every paper you might cite | Paste the model's bibliography into a manuscript |
| A temperature check on a public argument | A live model or a normal search | Treat it as a source |
| Repetitive file work on the teaching vault | A coding agent, after you have read what it can touch | Point it at a hospital directory |
| A drawing that is not a patient | An image model, labeled as a drawing | A generated picture standing in for a real case |
Paste this once
Put it at the top of the project, the custom instructions, or the file your agent reads every time. Then stop repeating yourself in every chat.
You are helping a surgeon prepare teaching material. Rules: - Assume the text is already de-identified. If you see a name, MRN, date of birth, face, voice, or a combination that could identify a patient, stop and say so. Do not continue. - Do not invent citations. If you are not sure a paper exists, say you are not sure. - Do not name anatomy that is not in the source. - Lead with the answer. Write for a fellow, in plain sentences. - This note is for one reader. Do not mention anyone else it might be for. - Flag uncertainty. Do not fill gaps.
The shape of a prompt
- The answer you need, in one sentence. “Cut this to 8 minutes.”
- Who it is for. A fellow, not “everyone.”
- What must not appear. Identifiers, invented papers, anatomy you did not describe, other people's names.
- The source text.
- Ask it to flag uncertainty instead of filling the gap.
Good prompts become notes in 40 AI / Prompts. Chat history is not a filing system. You will not find it on a Sunday.
How these models fail a surgeon
- Citations. They invent papers with real-looking authors. If you might say it out loud or put it in print, you open the paper.
- Anatomy. They will label a structure in a still they cannot see. Arrows are yours. The script names only what you named.
- Fluency. A smooth paragraph can skip the decision. Ask “what would make this wrong?” before you ask for polish.
- Agreement. They compliment a plan. Ask for the bailout and the complication, specifically.
- Other people's cases. Training text leaks in as if it were your case. If you don't remember it, it isn't yours.
- Authority. The model can hold an outline. You own the claim. A tool that decides for you is a different product, and it is not this one.
Agents, after the desk already works
A project brief is day one. An agent that edits files is later. When you get there: it may only see the teaching vault. The brief above is the first thing it reads. Give it one repeated job — chapter a video script, apply the case template — not “organize my life.”
Dictation is the agent most surgeons actually need. Speak the messy paragraph into the inbox. Clean it when you can see the screen. If the dictation service is in the cloud, the words have left the building. Don't dictate identifiers into it.
Mark this chapter
Saved on this browser. Not an account.
0 of 3 marked on this browser
Questions in this chapter
Can I paste an operative note into Claude or ChatGPT?
Short answer
Not if it still identifies a patient. Rewrite the lesson with the chart closed, then paste that.
- A header you deleted is not enough if the story, date, and surgeon still point at one person.
- Cloud dictation is the same door.
- When the text cannot leave the machine, use a local model or don't use a model.
Which AI should I pay for?
Short answer
One writing model you will open when you are tired. Keep a second only as a check when the first is fluently wrong.
- Put the project brief in whichever one you pick.
- Don't collect four logins and call it a workflow.
- Brand names change. The rule about what may leave the building does not.
Claude, ChatGPT, Gemini, or Grok — which one for teaching?
Short answer
Use the saved writing project for talks and scripts, a long-context app for PDF piles, and a live model only as a temperature check.
- Attach papers, then open every one you might cite.
- Grok is a natural fit for a public argument or a labeled drawing. It is not a quotation source.
- A coding agent touches the teaching vault only.
How do I write a prompt that doesn't waste the evening?
Short answer
Lead with the result, name the reader, state what must not appear, then paste the source and ask it to flag gaps.
- “Cut this to 8 minutes for a fellow” beats “make this better.”
- Save the prompt as a note once it works.
- Ask what would make the paragraph wrong before you ask for polish.
The model cited a paper I can't find. Now what?
Short answer
Drop the citation. If the claim matters, you find the paper yourself. A model bibliography is not a reference list.
- Search the title and the authors separately. Invented papers often mix real names.
- Don't “ask it to double-check” and then trust the second version.
- Write a decision note when it happens so you remember the fluency, not just the embarrassment.
Should I run a local model?
Short answer
Yes when the text cannot leave the laptop and a weaker draft is still useful. No as a badge, and not on weekend one.
- Ollama is the usual door.
- Expect worse long-form writing than the model you pay for.
- A local model does not fix a note you then paste into a cloud chat.
How do I dictate into the vault?
Short answer
Speak a messy paragraph into 00 Inbox. Clean it later on a screen. Cloud dictation means the words have left the building.
- Phone dictation or a desktop dictation app both work if you respect that rule.
- Don't dictate names, numbers, or room chatter.
- The inbox is for sentences, not for finished teaching notes.
What is a project brief, and do I need one?
Short answer
Yes. It is the short rule sheet pasted at the top of a project so you are not re-teaching the model every night.
- Use the brief in the AI chapter.
- It says: stop if you see an identifier, don't invent papers, don't invent anatomy, write for one reader, flag gaps.
- An agent that edits files is optional and comes after this, pointed only at the teaching vault.
How do I keep a model from inventing anatomy?
Short answer
You name every structure in the source. The model may tighten sentences. It may not add a label you did not write.
- Arrows on a still are yours.
- If it inserts a plane or a vessel, delete it.
- Say so in the project brief, not only in one chat.
Can I use AI to draft a video script?
Short answer
Yes, from your outline, on a de-identified description of what the frame shows. You record it, and you disclose a synthetic voice.
- Present tense. Structure first, then the move.
- Leave a quiet beat so the picture lands before the sentence.
- Read it aloud. A stumble means rewrite, not another take.