RPPA, Rheumatology Private Practice Alliance

Growth Summit 2026 · Omni Houston

The Sunday AI Workshop

Two hours, six modules, and two tracks: one for beginners and one for people already using AI. Hosted by Johnny Chan with Felix Lin, Xiao Zhang, and Andrew Eckstrom. Everything shown is free or nearly free, and every module ends with things you can do today.

Before Sunday (about 15 minutes)

Need help getting set up? Stop by the AI booth by Registration on Friday or Saturday. A coach will walk you through this list so you are ready on Sunday.

What you do not need

How the two tracks work

Beginner is each module's list of five things to do today. If you have never typed into a chatbot, start here. You will leave here today with something you can use tomorrow.

Intermediate is for people who already use AI every week, or who finish the beginner track early. Three builds per module that go one level deeper. Still no patient data, still free. If you are intermediate, congratulations, you are also a table captain on Sunday. Please help the two people nearest you before you start your own build.

The six modules

1AI Self-Defense for Private PracticeProtect

The same qualities that make AI useful, that it is cheap, fast, and sounds like a person, make it the best fraud tool ever built. Physicians are a high-value target, with a public name and photo, an NPI, an office that moves money on emailed instructions, and aging parents.

The three scenarios that may reach you

The voice that sounds like family. Sixty seconds of audio from a conference talk or a voicemail greeting is enough to clone anyone. The call goes to your spouse, your child, your parents, or your office manager. It is always urgent and it always needs money or a password.

The vendor who changed banks. A reply inside an existing email thread, in the vendor's real voice, with new wiring instructions. AI removed the typos that used to give these away.

Your face, selling something. A video of you endorsing a supplement. A telehealth site with your name and headshot. A wave of one-star reviews written overnight.

Urgency is the giveaway. A real emergency survives a hang-up and callback whereas a scam does not.

Five things to do today

  1. Pick a family safe word. One word, shared in person, never texted. Any call asking for money or secrets gets asked for it. Include your parents. If you have kids, turn on parental controls to limit screen time, filter content, and restrict chat rooms. Together in the room · 2 minutes
  2. Post the callback rule. No payment, wire, bank change, gift card, or password reset on a call or email alone. Someone calls back the number already on file. 10 minutes, one page, on the wall
  3. Lock your key accounts and your credit. Use a password manager. Turn on passkeys or a hardware key for email, bank, EHR, and payroll. Where those are not offered, use an authenticator app, not text codes. Turn on automatic updates for your computer, phone, and apps. Freeze credit at Equifax, Experian, and TransUnion. Also freeze ChexSystems, which blocks new bank accounts in your name. All free. 30 minutes, once
  4. Protect your own name. Google Alert on your name and your practice. Look up your NPI at npiregistry.cms.hhs.gov and check that the address is yours. If you bill Medicare, have whoever manages your PECOS login confirm the bank account has not changed. Bookmark the "report a fake account" page on Facebook, Instagram, LinkedIn, and Google so you can act within minutes. 15 minutes
  5. Write the practice AI rule. Which tools staff may use, that no patient information goes into any tool without a BAA, and what to do when someone slips. Have AI draft it, then read it. Then turn off model training on every AI account you use. In ChatGPT, go to Settings, then Data controls, and turn off "Improve the model for everyone." In Claude, go to Settings, then Privacy, and turn off "Help improve our AI models." This keeps your chats out of future models, but it does not make a tool safe for patient information. 10 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Draft a one-page AI acceptable-use policy for a small rheumatology practice. Cover which tools are approved, who our privacy officer is, the rule that no patient information enters any tool without a signed BAA, how staff report an accidental disclosure, and how every staff member signs it at hire and each year. Plain English, no legalese.

Say this, then hang up

To the caller: "I am going to hang up and call you back at the number I have. What is our safe word?"

To the vendor: "Thanks. I will confirm the new account by phone with the contact on file before anything changes."

To the patient asking about your video: "That is not me. Please do not buy anything, and thank you for telling us."

Use AI to fight back

Practice first with this fake email. It is safe: there are no links, and the address can never be a real website. One copy, one paste.

PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
Is this email a scam? List the warning signs and what to do. Subject: URGENT: Updated payment details before Friday From: "Dana Reyes, Accounts Receivable" <ar@infusion-supply-billing.example> Hi, Due to an audit at our bank, we have moved our payment account. Please update our banking details before Friday's payment run, or your next infusion supply order will be placed on hold. New account: First Coastal Bank, routing 000000000, account 0000000000. Please do not call our office line, it is down for maintenance. Reply to this email today to confirm the change. Thank you, Dana Reyes Accounts Receivable

Then use this one on real suspicious emails.

PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
Is this email or text a scam? List the warning signs and what to do. [PASTE THE MESSAGE OR ATTACH A SCREENSHOT. NO PATIENT INFORMATION]

Intermediate: one level deeper

Still no patient data in any tool.

  1. Run a tabletop drill with your office manager. "The doctor just emailed asking for a wire. What do we do?" Time how long the callback rule takes to kick in. If the answer is "we would have paid it," the rule is not posted where it needs to be.
  2. Set up a canary. Add a fake vendor to your contacts with a unique email address that exists nowhere else. If that address ever receives mail, your contact list has leaked and you know it the same day.
  3. Ask your malpractice carrier and any cyber insurance broker, in writing, whether you have cyber coverage, what the limit is, and whether social engineering and funds-transfer fraud are covered. Many malpractice policies include a small cyber add-on that excludes both. Save the answers with the policy.
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Draft a short, friendly, professional email from me to [MY MALPRACTICE CARRIER OR INSURANCE BROKER]. I am [MY NAME] at [PRACTICE NAME]. Ask them to answer these questions in writing: 1. Do I have any cyber coverage? Is it part of my malpractice policy or a separate policy? 2. What is the coverage limit, and what is the deductible? 3. Does it cover social engineering fraud, such as a fake vendor or fake doctor email that tricks my staff into sending a payment? 4. Does it cover funds-transfer fraud, where someone moves money out of the practice bank account? 5. If any of these is not covered, what would it cost to add, and what limit do you recommend for a practice our size? Ask them to reply by email so I have the answers on file. Keep it under 200 words. Do not guess or invent any details about my policy.
Why each defense works

The safe word. A voice clone reproduces sound, not shared memory. A safe word is a fact that was never typed anywhere an attacker could scrape.

The callback rule. Email is a message, not an identity. A callback to a number you already had forces the attacker to control a second channel they usually do not have.

Passkeys. A password can be typed into a fake page. A six-digit text code can be relayed in real time. A passkey is cryptographically bound to the real website, so a look-alike login page receives nothing. Authenticator apps are next best.

Credit freezes. No lender can open an account in your name until you unlock it. ChexSystems is the one most people miss. It covers new bank accounts, which is where stolen identities become cash.

Watching your name. You cannot stop a deepfake from being made. You can shorten the time it runs. Your NPI and DEA numbers are worth more than your credit card.

If it already happened

A scammer is on the line or on your computer: stop talking to them. If remote access software is running, end the session, then unplug the desktop, close the laptop lid, or pull the battery.

Money was wired: call your bank's fraud line immediately, ask for a recall, and file at ic3.gov the same day.

A clone or deepfake of you is circulating: report it as impersonation on the platform, save copies with URLs and dates, and notify your state medical board if it involves practice claims.

Credentials were phished: change the password from a different device, revoke active sessions, turn on passkeys, and check forwarding rules in your email settings, which attackers add to stay in.

Patient information went into a chatbot: do not panic. Delete the chat, write down what was shared and when, and tell your privacy officer that day. A short risk assessment decides whether it must be reported, and a small slip often ends with a note in the file.

A relative sent money to a voice clone: call their bank first, then the police report, then ic3.gov. While shame keeps people quiet, speed gets money back.

In every case: save every message and transaction record, and report it at reportfraud.ftc.gov and ic3.gov, plus identitytheft.gov if your identity was used.

Also on the horizon

Deepfake job applicants for remote billing roles, so verify ID in person for anyone who touches money or charts. AI scribes, so read before you sign and know whether your state requires disclosure. Cyber insurance exclusions, so ask your broker in writing.

Sources: FBI IC3 (ic3.gov), FTC consumer alerts on voice cloning, HHS OCR guidance on AI and business associates, NPPES, PECOS, NCMEC Take It Down.

2Front Office AutopilotOperate

Every private practice runs on a stack of unwritten knowledge: how faxes get triaged, what the front desk says when a patient is angry, how a new MA learns the ropes. Most of it lives in one employee's head. If she quits, that knowledge walks out the door with her. AI fixes this in an evening. You describe how a task gets done, in plain words, the way you would explain it to a new hire. AI turns that into a clear, step-by-step page you can print, laminate, and hand to anyone.

Where AI Can Help

The linchpin employee. One person knows how new patient intake actually works. Her training plan for a replacement is "shadow me for a month."

The 2019 photocopy. The gout diet handout has been copied so many times the header is gray, and it is written at a reading level half your patients cannot understand.

The 7am inbox. The office manager's first 90 minutes every day is triaging messages by hand, and the important one is always at the bottom.

Five things to do today

  1. Build one SOP (standard operating procedure). Fill in the recipe prompt's two blanks: the task you would hate to lose, and the job title of whoever does it. Together in the room · 10 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Write a step-by-step SOP for [TASK] at a small private rheumatology practice. It will be performed by [ROLE]. Include numbered steps, a checklist of supplies and system access needed, the three most common failure points and what to do about each, and when to escalate to the physician. Keep it to one page.

    Stuck picking a task? Try one of these: new patient intake, refill requests, closing out the day, onboarding a new MA, or infusion scheduling. No task in mind at all? Use ours below.

    ExampleReady to paste. Nothing to fill in.
    Write a step-by-step SOP for handling incoming referral faxes at a small private rheumatology practice. It will be performed by the front desk coordinator. Include numbered steps, a checklist of supplies and system access needed, the three most common failure points and what to do about each, and when to escalate to the physician. Keep it to one page.
  2. Ask for more versions. Stay in the same chat and send these one at a time, waiting for each answer. Together in the room · 5 minutes
    Follow-up 1Ready to paste. Nothing to fill in.
    Turn this into a one-page laminated checklist.
    Follow-up 2Ready to paste. Nothing to fill in.
    Write the 15-minute training script for a new hire.
    Follow-up 3Ready to paste. Swap Spanish for any language your patients speak.
    Now a version in Spanish.
  3. Rewrite one handout. Use any patient handout you already use. It is your document, not patient data. 5 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Rewrite this handout at a 6th grade reading level. Keep it to one page. [PASTE THE HANDOUT]

    No handout with you? Have it write a fresh one.

    ExampleReady to paste. Nothing to fill in.
    Write a one-page gout diet handout for patients at a 6th grade reading level. Include foods to enjoy, foods to limit, and three simple tips. Warm, plain English, no medical jargon.
  4. The week-sort. Dump your whole to-do list into the chat and ask it to sort into must do, should do, can wait, and tell you what you are forgetting. 5 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Here is everything on my plate this week: [LIST EVERYTHING ON YOUR PLATE]. Sort it into must do, should do, and can wait, and tell me what I am forgetting.

    Want to see it work first? Try this made-up week.

    ExampleReady to paste. Nothing to fill in.
    Here is everything on my plate this week: renew my DEA registration, call the billing company about a stack of denials, interview two MA candidates, finish my CME hours, fix the office printer, review the new lab contract, and plan my daughter's birthday party. Sort it into must do, should do, and can wait, and tell me what I am forgetting.
  5. Front desk scripts. Have AI write scripts for the three hardest calls your front desk takes: the angry patient, the same-day appointment demand, and the bill dispute. Ask for two versions of each, both warm but firm. Then have your front desk edit them, since they hear these calls every day. 10 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Write the phone script for the three hardest calls my front desk takes: the angry patient, the same-day appointment demand, and the bill dispute. Warm but firm, two versions each.

The recipe prompt

Intermediate: one level deeper

Still no patient data in any tool.

  1. The onboarding week. A day-by-day first-week checklist for a new MA or front desk hire, built from the SOPs you already made. Day one is badges and logins. Day five is shadowing the hardest task alone.
  2. The no-show letter series. Gentle first touch, firm second, final policy letter. Three letters, one prompt, then your office manager edits the tone.
  3. The website FAQ page. The fifteen questions your front desk answers by phone every week, written once, posted once, and the phone gets quieter.
PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
Here are my SOPs: [PASTE YOUR SOPS HERE]. Using them as the source, write a day-by-day first-week onboarding checklist for a new [ROLE] at a small rheumatology practice. Each day has three to five items, who signs off, and what "done" looks like.
More documents AI can draft for your office
  • A fee and billing explainer patients actually understand, at 6th grade reading level
  • Referral packet cover sheet: what a referring office should send with a new patient
  • Holiday and vacation coverage plan, written down for once
  • Job ads for open roles, plus interview question sets with a scoring rubric
  • Staff meeting agendas from a voice-note brain dump
Guardrails

You are the editor. Nothing patient-facing or staff-binding ships unread. Let the person who does the task edit the SOP. The draft is a conversation starter with your staff, not a decree from the chatbot.

3Vibe Coding PlaygroundPlay

The gap between "I have an idea for an app" and "it now exists" used to be a computer science degree. Now it only needs the audacity to ask. Building things you actually use, for productivity or fun, is the best way to learn AI.

Where AI Can Help

The idea you back-burnered. Some tracker, calculator, or family thing filed under "one day, when I have time to learn to code." That barrier to entry no longer exists.

The dashboard. A morning screen with the week's must-dos, one page, built by describing it. Not bought, not subscribed to, described.

The forgotten money. States hold unclaimed deposits, refunds, and final paychecks. One in seven people have some, and physicians move states more often than most.

Five things to do today

  1. Build your own one-page website. A birthday page, a family recipe page, a countdown to a trip. One prompt, one file. Steps to put it online are right below the prompts. Anyone with the link can see it, so leave off home addresses, phone numbers, and anything about patients. Together in the room · 20 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Build me a one-page website in a single HTML file. It is [WHAT IT IS, for example a birthday page for my mom, our family recipes, or a countdown to our vacation]. Include [THREE THINGS ON THE PAGE]. Make it warm, simple, and easy to read on a phone.

    No idea yet? Paste ours, then make it yours with the follow-up.

    ExampleReady to paste. Nothing to fill in.
    Build me a one-page website in a single HTML file. It is a birthday page for my mom. Include a big happy birthday headline, a short heartfelt message, and a section called five reasons we love you. Make it warm, simple, and easy to read on a phone.
    Follow-upTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Her name is [MOM'S NAME] and she loves [ONE THING SHE LOVES]. Add that, and use her favorite color, [COLOR].

    Put it online, free

    In Claude

    1Click your finished page so it opens on the right side of the screen.
    2Click Publish in the top right corner. If you see Share instead, click it and change “Only you” to Anyone with the link.
    3Copy the link and text it to anyone. They do not need a Claude account to open it.
    ?No Publish or Share button? Type: “Show this as an artifact so I can publish it.” The page opens on the right with the button. Still nothing? Download the file and follow the Netlify steps under ChatGPT.

    In ChatGPT

    1Type: “Give me this as a downloadable file named index.html.” Then download the file.
    2On a laptop, not a phone, go to app.netlify.com/drop and drag the file onto the page.
    3Netlify shows your link right away. Sign up free within an hour, or Netlify deletes the site.
  2. The unclaimed money sweep. On your own time, not in the room. Official state sites plus MissingMoney.com only. Search yourself, prior names, and your parents. Never pay a finder. The claim itself needs your SSN and ID, so the human files it directly with the state. 15 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    I have lived in [STATE, STATE, STATE]. Go to each state's official unclaimed property website and MissingMoney.com and search for [MY NAME AND PRIOR NAMES]. List every hit with the state, reported holder, and claim link. Do not submit any claims.

    Needs a browsing agent. This search only works with Claude in Chrome (paid Claude plans) or ChatGPT agent mode (paid ChatGPT plans), which can type into each state's search box. On a free plan, use this instead: “List the official unclaimed property website for each of [STATE, STATE, STATE], plus MissingMoney.com, with direct links.” Then search each one yourself. About 2 minutes per state.

  3. The promo code hunt. Next online purchase, let an AI browser agent (paid plans) hunt and test codes at checkout. Honest odds: often nothing, sometimes 15 percent. The human enters payment and places the order. 60 seconds
  4. Make something silly. "Build me a complete [GAME OR TOY] in a single HTML file" plus three sentences of description. Open the file, play it, show your kids. 15 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Build me a complete, playable [GAME OR TOY] in a single HTML file with no external dependencies. [THREE SENTENCES DESCRIBING IT.] Make it work on a phone in portrait mode with touch controls.

    No idea? Try this one.

    ExampleReady to paste. Nothing to fill in.
    Build me a complete, playable whack-a-mole game in a single HTML file with no external dependencies. Moles pop up from nine holes and you tap them before they hide. It speeds up every ten points and keeps a high score. Make it work on a phone in portrait mode with touch controls.
  5. The meal planner. Not the flashiest prompt here, but the one you will still be using months from now. 2 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    You are my meal planner. Family of [NUMBER OF PEOPLE], we like [CUISINES YOU LIKE], [LIMITS, LIKE BUDGET, ALLERGIES, OR 30 MINUTES MAX]. Give me a week of dinners and one grocery list.

Intermediate: one level deeper

Still zero patient data, still toys until they are not.

  1. The comparison shopper. Paste requirements for a big purchase (a car, an infusion chair, a phone system) and get a ranked table with the tradeoffs stated, not a listicle.
  2. Practice-adjacent toys. A waiting room trivia screen, an office holiday raffle picker, a "guess the joint" quiz for the lobby TV. Still zero PHI.
  3. The graduation rule. When a toy becomes a tool people rely on, learn hosting, backups, and logins before anyone else depends on it. Ask the AI to explain each one to you like a new hire, then do them.
PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
I am choosing between [OPTIONS] for [PURPOSE]. My requirements, in priority order: [YOUR REQUIREMENTS]. Build a comparison table, score each option against each requirement, state the tradeoffs plainly, and tell me which one you would pick and why. Flag anything you are unsure about.
4The Referral EngineGrow

A private rheumatology practice lives on referrals. PCPs send patients to the specialist they remember, trust, and hear back from. The problem was never that specialists do not care. It is that personal touch does not scale past a 12-hour clinic day. AI collapsed that. A personalized video or a warm specific note to every top referrer, every month, all of it now costs just a few minutes and a small handful of pocket change.

Where AI Can Help

The referral that moved. A PCP who sent you patients for years quietly started sending them across town. No fight. Someone else was simply more present.

The unwritten thank-you. A colleague sent you a genuinely interesting case. You meant to call. It is now three weeks later.

The black hole. The number one complaint referring physicians have about specialists is silence.

Start with your referring doctors. A thank-you to a colleague contains no PHI and needs no patient consent. Keep it that way: "thanks for that complex case last week" is warm enough. Never include a patient's name or any details.

Five things to do today

  1. Draft five real thank-yous. Five notes, for five real colleagues, edited in your voice, ready to send Monday. Start here today. 10 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Draft a warm, specific thank-you note to Dr. [LAST NAME], a referring physician who just sent me [BRIEF, NO-NAME DESCRIPTION OF THE CASE]. Two paragraphs, no flattery, sounds like a busy human who appreciated the referral. Mention that I will keep them posted. Do not include any patient details.
  2. Build the top-five list. List the five doctors who refer to you most, in a note on your phone. Next to each name, write the last time you reached out personally, not just a consult letter. The list tells you who to thank next. The videos and notes are how you do it. 10 minutes
  3. Make it monthly. Send each of your top five referrers one personal note or video a month. Put the first one on the calendar. 2 minutes
  4. Make your avatar at home, not in the room. Sign up for an AI video service like HeyGen. Somewhere quiet, record yourself reading its sample script into your phone camera for about 60 seconds. That is all it needs to make a talking video of you. 15 minutes, mostly waiting
  5. Send yourself the test video. Type a three-sentence thank-you to a made-up colleague, have your avatar say it, and watch it. Seeing yourself on screen is what makes this click. When you send real ones, add a light note so nobody mistakes it for live video, like "made this with my AI clone, seemed more fun than a fax." 10 minutes

Intermediate: one level deeper

Still no patient data in any tool, and nothing patient-facing without checking your state's AI disclosure rules.

  1. The quarterly practice update video. Two minutes, "here is what is new at our practice," sent to every referring office. One recording session, evergreen for three months.
  2. Welcome-to-town notes for newly arrived PCPs before they have referral habits at all. The highest-leverage note in this module.
  3. The referral-back letter template for the other direction, when you return a patient to primary care with a clear plan. Template only, filled in on office systems.
PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
Write a two-minute spoken script for a quarterly update video from a rheumatology practice to its referring primary care offices. Cover one new service, one scheduling improvement, and one thing we want referrers to know before they send a patient. Warm, specific, no jargon, no patient details.
More ideas
  • Lunch-and-learn decks for PCP offices: "what to work up before you refer, and what to just send me"
  • Holiday cards at scale, the original use case
  • Staff appreciation videos, the same trick pointed at your own team
Guardrails

You are uploading your face and voice to a vendor. Use a strong unique password, know the deletion policy before signing up, and treat that account like a bank login.

5The Study MachineLearn

Physicians are students for life, but new papers pile up faster than anyone can read them. AI turns any paper into the formats that actually stick: audio for the commute, flashcards, or a two-page summary. Upload the PDF and type one simple sentence.

Where AI Can Help

The tab. A guideline update has been open in a browser tab for three weeks. It has been there so long you have learned to ignore it.

The photo. A conference slide you photographed, fully meaning to study it later. It is still buried in your camera roll.

The night before journal club. It is 11pm, you are presenting a dense paper to your colleagues tomorrow, and you are summarizing it by hand like it is residency again.

Five things to do today

  1. Turn one guideline into a podcast. Upload the PDF to Gemini Notebook (Google's free AI notebook, formerly NotebookLM) and click Audio Overview. In a few minutes it makes an engaging podcast-style summary of the paper. Download it to your phone and play it on tomorrow's commute. 10 minutes
  2. Turn one journal article into flashcards. Paste the article into Claude or ChatGPT with the prompt below. 5 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Make ten flashcards from this paper. Question on one line, answer beneath, focused on what changes clinical decisions. [PASTE THE PAPER HERE]
  3. See only what changed. Give AI the new guideline and the previous version, and ask for just the changes. Reading only what changed beats rereading the whole document. 5 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Here is the new guideline and the prior version. List only what changed, why, and what I should do differently Monday. [PASTE BOTH VERSIONS HERE]
  4. Build the journal club one-pager. Ask AI for a one-page summary of the paper: study design, endpoints, effect sizes, the limitations the authors downplay, and three discussion questions. Then edit it yourself. You present better from a page you fixed. 10 minutes
    PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
    Summarize this paper for a journal club of rheumatologists: design, endpoints, effect sizes, limitations the authors underplay, and three discussion questions. [PASTE THE PAPER HERE]
  5. Start your personal library. Make one folder for each topic you follow (in Claude, a Project) and drop in every paper you read. By month three you can ask "what do my papers say about JAK inhibitor safety?" and get an answer drawn only from papers you chose, not the open internet. 5 minutes to set up

Intermediate: one level deeper

Still papers in, patients never.

  1. Board and MOC prep. Your personal library becomes a personal question bank generator. Ask for ten board-style questions on a topic, grounded only in the papers in the folder, with the source cited for each answer.
  2. The skeptic drill. Paste a too-good industry slide deck and ask "what would a methodologist attack first." Do it before the rep's next visit.
  3. The conference debrief. Slide photos plus voice memos in, one organized brief out, within 48 hours of flying home while it still matters. Start with Houston.
PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
You are a skeptical methodologist. Here is a slide deck summarizing a clinical trial. List the five things you would attack first: design choices, endpoints, comparators, subgroup claims, and anything the slides imply but do not show. Be specific. [PASTE THE SLIDE TEXT HERE]
More ideas
  • Teaching residents: "write five quiz questions on this topic, then grade my resident's answers against the source"
  • Statistics on demand: "explain the difference between this hazard ratio and what my patient will actually experience, with a worked example"
Guardrails

Verify before you cite. AI summarizes documents you gave it well, and invents citations you did not give it with total confidence. Any reference that leaves your hands gets checked against the source.

Read the original before you change care. A summary can leave out details that matter, like who was studied, the dose, or the exceptions. Before a summary changes how you treat a patient, read that section of the paper yourself.

Keep your own reading skills sharp. If AI summarizes everything for you, you can slowly lose the habit of reading papers closely. Once a week, read one paper the old way, start to finish, before you ask AI about it.

6One Bot, One JobDelegate

A chatbot answers when you ask, whereas an agent does a job. You already manage human agents, like an MA who rooms patients, a biller who works denials, a front desk that confirms tomorrow's appointments. Each one has a single job, a thing to produce, and a rule for when to come get the doctor. An AI agent is the same shape and everything here is free or nearly free and can quickly and easily be set up today.

Where AI Can Help

The Monday review pile. Online reviews of your doctors and staff pile up all week on Google and Yelp. Nobody reads them until one goes tragically viral, and by then the staff member it names has heard about it from a patient.

The thank-you that never gets written. You meant to thank the PCP who sent you three patients last month. It is now the month after.

The person who does everything. Your best front desk person also answers phones, chases billing, and orders supplies. Each job gets done well enough, none of them great, and when she is out, all of it stops. A bot works the same way. One bot, one job is how you get one you can trust and stop thinking about.

The three parts of any agent

  1. The job. One sentence, including where it looks. "Every Monday at 7am, summarize last week's Google reviews for my practice."
  2. What it produces. A one-paragraph email, a row in a spreadsheet, a draft reply, a document.
  3. When it stops and asks. "If a review names a staff member or mentions a safety concern, do not summarize, flag it to me."

Three rules

One job. No money, no patient data, nothing sent to patients. Agents draft, humans send. Every agent has a boss, and you read what it produces for two weeks before you trust it.

Five agents to build today

Each one is PHI-free and takes under 15 minutes. Write the three parts on paper first. Everyone can do that. Then run it as a saved prompt by hand each week. Move it to a scheduled task (Claude or ChatGPT paid tier) once the output has been right four weeks in a row.

  1. The review reader. Every Monday, summarize last week's Google and Yelp reviews for the practice. Flag any that name staff or mention safety. Agent output is one paragraph in your inbox.
  2. The referral thank-you drafter. Every Friday, give your agent the list of referring doctors for the week. It will draft one warm, specific note per name. It gives you drafts to approve and send.
  3. The inbox triage bot. Every morning, it reads one email folder you pick and sorts them into three short lists: needs you today, can wait, and ignore.
  4. The meeting prep bot. The night before any calendar event tagged "meeting," pull the last three emails with that person and write a five-line brief. It is in your inbox by 7pm.
  5. The website watcher. Once a week, it checks two or three web pages you care about, like your state medical board's news page or the American College of Rheumatology's guidelines page, and emails you a short note in plain English on what changed. If nothing changed, the note says so.
PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
I want to hand off one job to AI. The job: [what it does, when, and where it looks, for example, every Monday read my Google reviews]. What I want back: [for example, one paragraph in my inbox]. Stop and ask me if: [for example, a review names a staff member]. Write me step-by-step instructions I can reuse every week, the way I would train a new employee on day one. Then tell me if I should set this up as a scheduled task in Claude or ChatGPT, or just paste it in myself each week.

No job in mind? Start with this one.

ExampleReady to paste. Nothing to fill in.
I want to hand off one job to AI. The job: every Monday, read my practice's new Google reviews. What I want back: one paragraph in my inbox with the overall mood, anything to praise a staff member for, and a short, generic draft reply to each review that never confirms the reviewer is a patient. Stop and ask me if: a review names a staff member, mentions a medical complaint, or sounds like a patient in distress. Write me step-by-step instructions I can reuse every week, the way I would train a new employee on day one. Then tell me if I should set this up as a scheduled task in Claude or ChatGPT, or just paste it in myself each week.

Intermediate: one level deeper

Still no patient data, still agents draft and humans send.

  1. Move one agent from pasting it in yourself to a scheduled task after four weeks of correct output. Write down the date you promoted it and the first time it was wrong.
  2. Write the boss-check agent. Once a week, it reads the other agents' outputs and tells you which one drifted. The boss gets a boss.
PromptTap Copy, paste into ChatGPT or Claude, fill in the [BRACKETS]. No need to read it.
Here are the outputs from my agents this week, labeled by agent name. For each one, tell me whether it followed its three-part job description [PASTE EACH AGENT'S THREE PARTS HERE], whether its output drifted from last week's, and whether anything should have been escalated to me but was not. Short, one paragraph per agent.
Why the three parts work

Why one job. An agent with one job has one place to fail, so you can see the failure. An agent with five jobs fails somewhere in the middle and you find out three weeks later.

Why "stops and asks" matters most. The escalation rule is what makes an agent safe to leave alone. Write the rule before the job. If you cannot think of one, the job is not ready to delegate.

Why agents draft and humans send. An agent that sends can send the wrong thing to the wrong patient at 3am. An agent that drafts costs you a click and never a complaint.

Where agents run today. A saved prompt you run by hand each week, which is where most people should start. Then a scheduled task in Claude or ChatGPT (paid tier) once the output has been right four weeks in a row.

What an agent is not. Not a second brain, not a digital twin, not an employee replacement. A very fast, very literal new hire who reads only what you point at. That literalness is the feature.

No patient information goes into any AI chatbot. The only exception is a tool where your practice has signed a business associate agreement (BAA), the HIPAA contract that makes a vendor legally responsible for your patients' data, the same one you already have with your EHR and your billing company. If you do not know whether you have one, you do not. Everything on this page needs zero patient data.

After the workshop

Keep learning after Houston with one module a month. Each month, pick a module, complete the five items on its Beginner tab, then try the builds on its Intermediate tab.

Print versions

Two-page handouts for each module, print double sided. Links go live once the handouts are finalized.

This page is general information, not legal, financial, or compliance advice. Coaches: Johnny Chan, Felix Lin, Xiao Zhang, Andrew Eckstrom. Stay in touch: johnny@aiforprivatepractice.com
Copied. Now paste it into ChatGPT or Claude.