Where to start with AI when you’re a busy physician

Last week, after a Teams meeting ran short, a colleague stayed on the line. He is a practicing clinician and a hospital leader. He had been trying to use AI to analyze some hospital finance data, looking for areas of margin he could improve. He told me the output was not meeting his expectations and the whole thing felt like a fight. More work than quality.

I asked to see his prompt. It was one line.

“Help me to identify areas of the margin data we can improve on.”

That sentence is the problem most physicians have with AI. Not time. Not technology. The prompt.

If you have ever opened ChatGPT or Gemini, typed a short ask, gotten a generic answer back, and quietly decided AI was overrated, this post is for you. There is nothing wrong with you. You did not get a bad tool. You got a tool that wrote you a generic answer because you gave it a generic ask. The fix takes about thirty minutes.

A small confession before the recipe

I am a practicing anesthesiologist. I spent the past year stress-testing AI across the entire spectrum. Consumer chat tools (Gemini, Claude, ChatGPT). Image generation (Nano Banana 2). Voice (ElevenLabs, WhisprFlow). Search-style AI (Perplexity). Document tools (NotebookLM). Coding assistants (Claude Code, Codex, Copilot 365, Google Antigravity, the Claude and Gemini integrations in Xcode). Direct API access through Gemini and Claude. A full self-hosted stack of my own called OpenClaw, running multiple local models on Ollama and deployable on a free Oracle VPS.

Some of it worked. Plenty of it did not. None of it was clinical.

The first tool I ever tried that actually changed something was Gemini, planning an international family trip. The work I would normally do across several hours of browser tabs and color-coded spreadsheets, it did in ten minutes. That was the moment I stopped asking whether AI was useful and started asking which parts of life it could absorb so I had more brain left over for the parts that need me.

The tool that took me the longest to add was Claude. It can feel intimidating, and the pricing can feel high. Once I understood how the pieces work together (Claude chat for fast questions, Claude Cowork projects for ongoing work, Claude Code for anything programmatic), the value clicked. More on what I mean by “stack” in a later post.

Here is the line I find myself coming back to when a colleague asks where to start.

It is not as complex as social media and the news make it sound. Stick to the basics. Do not dip your toes into too many AI ecosystems at once.

The rest of this post is how I would walk a physician colleague through their first thirty minutes if they handed me their laptop and said “show me.”

A quick note before the steps. If you are not currently a Gmail user, Google AI Pro is still the best bang for your buck I have found, and what follows applies. If you would rather look at other setups (Microsoft 365 with Copilot, Claude, or ChatGPT as your daily driver), the steps are essentially the same. Pay for one ecosystem. Install the apps everywhere. Configure memory plus integration plus standing instructions. Then write a real prompt with context. The pattern is identical across ecosystems. Only the names of the buttons change.

Step 0 (before any prompt). Use the AI you already half-have.

If you use Gmail, your starting move is to upgrade your existing Google account to Google AI Pro at one.google.com/about/google-ai-plans. It is twenty dollars a month. You get a large bump in cloud storage, full Gemini Pro features (essentially unlimited access to a frontier-level model), and (this is the part people miss) deep integration with Gmail, Drive, Docs, Sheets, and Calendar.

This is the unlock most physicians do not realize they need. The free tier of any AI tool is a chatbox sitting in isolation. The paid tier with Workspace integration is an AI that already knows your email, your calendar, your documents. The first time you ask it to do something that touches your actual life, you stop having to copy and paste your context into the prompt. The context is already there.

Step 1. Install Gemini everywhere. Configure it once.

Before you write a single real prompt, do the boring infrastructure work.

1. Download the Gemini app on your Mac (or PC), on your phone, and as a Chrome extension. The point is that the same AI can hear you whether you are at your desk, in the car, or between cases.
2. Open gemini.google.com and go to Settings → Personal Intelligence.
3. Turn on memory.
4. Connect at least Google Workspace, so Gemini can see Gmail, Drive, Docs, Sheets, and Calendar.
5. Create a short set of standing instructions. Things you always want known. Your writing style. Your punctuation preferences (mine is “no em-dashes, ever”). Things you always want done. Things you never want done. Keep it short. Five or six lines is enough to start.

This setup pattern is the same in every major AI ecosystem. Once you have done it in Gemini, you can do it in Claude (Projects + custom instructions) and ChatGPT (Memory + custom instructions) in about three minutes each.

Step 2. The first real prompt.

This is the moment most people botch. They open the box and type three words.

The first prompt I would suggest is one that solves a real problem, has a clear “done” state, and runs forever once it works. Type something like this:

“Help me to create a daily task to run that provides me an update on my local news every day (including important events from the past 24 hours) at 5:30 AM.”

A few things to notice about that prompt. It names the cadence (daily). It names the time (5:30 AM). It names the scope (local news, important events from the past 24 hours). It is specific enough that Gemini can either schedule it as a recurring task or, if scheduling is not available in your region or plan yet, give you a one-tap workflow you can run yourself each morning.

The next morning, when the briefing appears, something quiet happens in your head. AI stops being a chatbot and starts being a small piece of staff.

Step 3. The second prompt. Day two.

Now that the briefing is running, try the one that pays back the most in the first week. Open Gemini and type something like:

“Evaluate emails from the last 7 days that were sent to me and required a response, but I have yet to respond to. Sort this list by priority, including deadlines, bill payments, or sense of urgency.”

This is the prompt that makes physicians stop saying “AI is interesting” and start saying “AI is useful.” Because you set up the Workspace integration in Step 1, Gemini already has access to your inbox. It does not need you to copy and paste anything. You will get back a real, prioritized list of emails you actually owe a response to.

Two prompts. Two real wins. You are forty-eight hours in.

Why the colleague’s prompt failed (and what mine did differently)

Back to the hospital leader on the Teams call. We did two things to his prompt. Neither was complicated.

First, we changed the model. The hospital provides Copilot 365, which in Auto mode usually routes to ChatGPT under the hood. He did not realize he could explicitly select Claude Opus within the same Copilot ecosystem. Different models, different strengths. For dense, structured analytical work, Claude Opus often produces a noticeably better first pass.

Second, we restructured the prompt itself. The original was one short sentence with no context, no constraints, and no requested format. The rewrite followed a prompt framework called BRIEF. There is a sibling framework called COSTAR. Both give you a small checklist (background, role, expected output, examples, format, and so on) that turns a vague ask into a structured one. There is also a useful shortcut when you are short on brain power: have an AI generate a high-quality prompt for you to then run on another AI. I will write a dedicated post on BRIEF, COSTAR, and the prompt-the-prompter trick. For now, the takeaway is small.

A short prompt produces a short, generic answer. A structured prompt produces a useful one. The fight you are having with the AI is almost always a fight with the prompt.

A real failure I want you to learn from

I do not want to leave the impression that I have only had wins. I have not.

The most expensive lesson of my year was OpenClaw. OpenClaw is my self-hosted stack, formerly named moltbot. It runs Ollama as the model runner with Open WebUI as the front end, plus a voice and personalization layer on top, plus an autonomous-agent loop that is supposed to keep working on a task without me. I spent several weeks setting it up on a free Oracle VPS. I had been told this kind of setup would change my life because it would “do, do, do until the task was complete.”

What actually happened was two things. First, I almost made the classic mistake of not locking down my API keys, which can quietly burn through usage limits and produce a real bill (I had a small dollar cap set, so I escaped with a cheap lesson; the full API-key safety story will get its own post). Second, the autonomous loop would either churn for hours and then fail, or finish and produce nonsense.

The lesson I took, and the one I want you to take, is this.

Be careful how much you trust an AI to be a good steward of resources. Money, compute, time, and the work itself. It usually will not be.

That is the line that should sit behind every decision you make in the first month. Use AI to draft and accelerate. Use a human (you) to verify and decide. There is a separate failure story from this same season, where an AI in autonomous coding mode deleted an entire app database while I was at work because I wanted it to keep working in the background. I will tell that one in a dedicated post on autonomous-mode safety.

The three traps in week one

Once your news briefing is firing and your inbox triage is humming, three traps tend to bite.

1. Chasing more without understanding the process. Trying to get more and more out of the tool without slowing down to learn how the tool actually works.
2. Expecting every output to be perfect. They are not. They are first drafts. Two or three rounds of plain-language refinement is normal, not a sign you are bad at this.
3. Trusting authoritative-sounding output without verifying. The output will sound confident even when it is wrong. The rule is simple. Trust but verify, especially for anything involving numbers, names, or facts where being wrong actually matters.

These three traps are physician-shaped. We are trained to want complete, correct, authoritative answers in the first pass. AI does not work like that. It works like a fast, confident, slightly over-eager intern. You are still the attending.

On privacy and the hard line

The most important sentence in this entire post is the one I will write twice.

Never put patient information into a non-HIPAA-compliant system.

Never put patient information into a non-HIPAA-compliant system.

That covers Gemini, ChatGPT, Claude in its consumer form, Copilot in its consumer form, and any self-hosted setup you have not explicitly hardened. The non-clinical guardrail of this site is not a marketing tagline. It is a real liability line. Stay on the right side of it.

A few practical privacy notes for the personal side.

• When you pay for an AI ecosystem (Gemini, Claude, ChatGPT), you can usually opt out of having your data used to train future models. Do it. The toggle is in the settings.
• Even with training opted out, treat the system as if a future visitor to your account could read your full chat history. Do not paste your most sensitive material in (full financial statements with account numbers, photographs of legal documents, anything you would not be comfortable surfacing in a leak). Use the AI to draft, not to store.
• Be especially careful with anything in autonomous mode. Autonomous agents are vulnerable to prompt injection and can execute code on your behalf. The further the agent gets from your direct attention, the higher the cost of something going wrong.

Two questions that come up next

Will using AI make me lazy or worse at thinking? It can. It will, if you use it as a substitute for thought. But if you use it to absorb the “extra” (the busywork, the email triage, the calendar wrangling, the meal plan, the rebooking of the kid’s flight), you free your actual brain for the work that genuinely needs you. The honest physician version is, use AI to reduce the extra, not to outsource the thinking.

Am I behind? Not at all. This may be the fastest-growing field, and the fastest-spreading consumer product, in modern history. Most people will not ever reach true expert level. That is fine. The goal is not to become an AI expert. The goal is to reach the basic level that improves your life without stressing you for being inadequate.

You are not behind. You are busy. There is a difference.

The point

The biggest mistake physicians make with AI is not picking the wrong tool. It is picking the wrong starting move. You do not need a weekend. You do not need a new app. You need thirty minutes, a Google AI Pro upgrade, a configured Gemini, and two real prompts.

Do that this week. Then come back. I will be here, writing about the stack, the prompts, the failures, and the parts of life that medicine does not give back.

If any of this got too complex

Leave a comment below and tell me where you got stuck. I will write a simpler version of that section, in plain language, no jargon. Peer to peer is the whole point of this site, which means the conversation goes both ways. No question is too basic. Most of the questions I have wished I could ask out loud over the past year were the basic ones.

Strictly non-clinical. Nothing on this site is medical advice. I do not post about patient care.