The AI Content Flywheel: How I Build an Audience as a Surgeon Without Sounding Like a Robot

You're Prompting Wrong: A Claude Cowork session with Dr. Christian Péan

Thank you to everyone who tuned into my live video! Join me for my next live video in the app. Prompt With CARE What watching someone else prompt taught me about the skill we all think is basic—plus two prompt packs that set up your Claude Cowork in one afternoon Yesterday I asked a student to sit at my computer and prompt Claude Cowork to build a presentation. He typed a single sentence, hit enter, and waited. Watching him, I realized two things: the conversational interface with these tools is a genuine skill, and I’d been doing it so long I’d forgotten it had to be learned. I teach people AI tools constantly—I had never once sat down and watched someone else prompt. It was the most instructional thing I’ve done in months. So this piece is the session I ran afterward, in writing. One framework, one habit loop, and—because readers keep telling me prompt packs are the most useful thing I publish—two packs you can paste into Claude Cowork today: a Cold Start Interview that sets up your workspace from scratch, and five Workflow Interviews that teach it your voice, your week, and your standards. Nothing here is specific to medicine. If you have a job and a desk, this applies. Housekeeping: I don’t type my prompts—I speak them. You will never type as fast as you talk, and the more descriptive your language, the better your output. I use Wispr Flow [https://ref.wisprflow.ai/christian-pean-wi5k] (function + spacebar, then just talk—you can literally whisper). If you want to try it, downloading through my affiliate link [https://ref.wisprflow.ai/christian-pean-wi5k] supports this newsletter. The Frappuccino Problem A close friend of mine has a habit that drives me a little crazy. He’ll say, “I don’t like the way AI writes X.” or he’ll say “AI never makes my images or powerpoints right”. And every time, I think: it’s not that you don’t like what AI makes. It’s that you don’t like what AI makes for you…because you’re giving it poor instructions. Think of these tools as a very smart intern. Maybe a PhD-level Harvard graduate of an intern. If I tell that intern “go get me coffee,” and they come back with a Frappuccino when I wanted an Americano with oat milk, that’s on me. The intern did their best with what I gave them. A language model is the same: every detail you leave out, it fills in with the most average assumption available. Vague prompts don’t produce bad output. They produce median output, which for any real professional task is the same thing. The fix is a framework I call prompting with CARE: C — Context. Who you are and what situation this output lives in. A — Audience. Who actually reads, sees, or uses this. R — Role. Who the AI should be—”you are a patient educator,” “you are a skeptical CFO”—so it pulls the right body of knowledge into the task. E — End product. The exact artifact you want: format, length, tone, what it should be usable for the moment it’s done. In my live session I demonstrated this with two prompts about the same topic. “Tell me about hip fractures” got me a long, jargon-heavy wall of text. The CARE version: “I’m a trauma surgeon, fewer than a third of my patients ever start bone-protecting treatment after a fracture, you’re a patient educator, write for families at a sixth-grade reading level, one-page handout, warm and non-alarmist, end with three questions to ask your doctor” This prompt instead produced something I could physically hand to a patient that afternoon. Same model. Same topic. The difference was entirely in the briefing. Your First Output Is the Input The second principle, and the one that separates people who get compounding value from people who quit: the first output is not the final output. It’s the input for the next cycle. You will be very frustrated if you try to one-shot these tools. The entire power of a large language model is the compression of iteration cycles: you get unlimited at-bats, and each swing takes seconds. When my bone-health handout came back and I didn’t love it, I didn’t start over. I said what needed to change. Make it Duke-branded, add interactive components, also give me a PowerPoint I can click through. And the output became the raw material for something better. Later I dragged that same HTML into a fresh Cowork task and rebranded it for RevelAi in one prompt. Output becomes input. That’s the loop. Two force multipliers on that loop. First, speak instead of typing. Being verbose is usually a weakness of mine; it turns out it’s a superpower with these models, because descriptive, rambling, context-rich instruction is exactly what they reward. You’re not on a timer, and nobody’s grading your dictation. Second, meta-prompting [https://techysurgeon.substack.com/p/stop-writing-prompts-start-building?r=3adnor]: when the end product demands a better description than you can produce—a detailed image prompt, say—don’t write the prompt yourself. Ask the tool to write the best-practice prompt for you, then carry it wherever you need it. The model is better at describing what the model needs than you are. The Step Most People Never Take Everything above makes a single task go well. The better approach is making sure you never start from zero again. This is what I think of as building sustainable context systems. The low-hanging fruit is Projects: a folder where your chats, files, and outputs accumulate so the tool carries the thread across sessions. I run my academic promotion tracker, cap table, and investor touchpoints this way connected with Live Artifacts. Learn more about Scheduled Tasks and Live Artifacts Below: But the foundation underneath all of it—the thing I do with every person I onboard one-on-one—is to have Claude interview you. Instead of laboring to write the perfect context document about yourself, you flip it: make the model ask you questions, one at a time, until it understands your role, your projects, your people, and your standards. Answering questions is cognitively cheap. Composing instructions is expensive. And in Cowork, the answers don’t evaporate when the chat ends—they become standing infrastructure, so every future task starts with the CARE already half-written. The question is what the interview should cover. That’s the part most people improvise badly. Below the line are the exact packs I give people in one-on-ones—six prompts that stand up your workspace from nothing, and five that teach it your recurring workflows. The two prompt packs below are for paid subscribers. Pack One stands up a Claude Cowork workspace from scratch in about thirty minutes—it’s the same starter kit I use when I onboard people one-on-one. Pack Two teaches your Cowork your writing voice, your calendar, and your definition of done. Copy, paste, answer the questions out loud and you’re all set. As a bonus, in the paid section… It seems like many people are a fan of the one-shot video note [https://substack.com/@techysurgeon/note/c-271385379?r=3adnor&utm_source=notes-share-action&utm_medium=web] that I wrote recently. I also included a video of me installing the Higgsfield MCP [https://higgsfield.ai/mcp]. Techy Surgeon is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Prompt Pack One: The Cold Start Interview Run these in order, in Claude Cowork, ideally in one sitting. Each prompt instructs Claude to interview you and then write what it learns to memory—so say “save this” when the summary looks right. Total time: about thirty minutes. Less if you dictate. 1 — The Role Brief You’re setting up as my long-term work assistant, and right now you know nothing about me. Interview me, one question at a time, until you understand: what my job actually is day to day (not the title—the verbs), who I’m accountable to, what I produce, and what a great week versus a bad week looks like. Don’t accept vague answers—push for specifics. When you’re confident, write a one-page working profile of me, show it to me for corrections, and save it to memory. 2 — The Project Map Interview me about everything I’m currently working on—projects, initiatives, recurring responsibilities, and the things that are stalled but still mine. One question at a time. For each, capture: what it is, current status, who else is involved, the next milestone, and what I’m most worried about. Then organize it into a project map ordered by what deserves my attention, show me, and save it to memory. 3 — The People Directory I’m going to mention names, nicknames, and abbreviations constantly, and I never want to explain them twice. Interview me about the people and organizations in my work orbit: who they are, their role relative to me, how I refer to them in shorthand, and anything you should know before drafting something they’ll read. One question at a time, and prompt me with categories I might forget—boss, reports, clients, vendors, collaborators, the person I always forward things to. Save the directory to memory. 4 — The Preference Card Interview me about my output preferences so you stop guessing. Cover: how long answers should be by default, bullet points versus prose, how formal my drafts should run, formatting pet peeves, words and phrases I’d never use, and how I want you to behave when you’re uncertain—guess, ask, or give options. Get concrete by showing me two short sample paragraphs in different styles and asking which is closer. Save the result as my standing preference card. 5 — The Definition of Done For each type of deliverable I produce regularly—you have the list from my project map—interview me about what “finished” means. What does a ready-to-send email have that a draft doesn’t? What makes a document ready for my boss versus ready for a client? Build me a short rubric per deliverable type, confirm it with me, and save it. From now on, check your own work against these rubrics before showing me anything. 6 — The Standing Orders Last one. Interview me about the rules of engagement: things you should always do without asking, things you should never do without asking, topics where I want pushback versus execution, and how proactive to be when you notice something off in my files or messages. Compress what you learn into a set of standing orders, read them back to me, and save them. Then summarize everything you now know about me from this whole setup in ten bullets—let’s see how well you listened. Prompt Pack Two: The Workflow Interviews Run these as needed, after the Cold Start. Each targets one recurring workflow and ends with a reusable asset. 1 — The Voice Extraction I’m going to paste three pieces of writing that sound like me. Read them, then interview me—one question at a time—about the choices you noticed: sentence rhythm, how I open, how I handle disagreement, what I do instead of corporate phrasing. Challenge me where my samples contradict my self-description; trust the samples. Then write a style guide for drafting as me, test it by rewriting one paragraph of your own in my voice, and ask me to grade it. Save the final guide to memory. 2 — The Weekly Rhythm Interview me about the shape of my typical week: recurring meetings and what I need before each, deliverables on a cycle, the days that are always crunched, and the prep work I always do in a rush at the last minute. Then propose three things you could take over on a schedule—drafts, briefings, summaries—with the day and time you’d deliver each. I’ll pick. Save the rhythm to memory and set up the scheduled tasks I approve. 3 — The Triage Spec I want you to be able to look at my inbox or message backlog and know what matters. Interview me about how I triage: which senders always get same-day replies, what “urgent” actually means in my world, what I can safely ignore, and what I’m always afraid of missing. Turn my answers into a written triage rubric with named tiers, confirm it with me, and save it. Then, whenever I ask you to review my messages, apply it and show your sorting. 4 — The Decision Brief When I bring you a decision, I don’t want a list of considerations—I want it framed for a verdict. Interview me about how I like decisions presented: how many options, how much detail per option, whether I want your recommendation up front or last, and what evidence earns my trust. Build my decision-brief template, run a real decision I’m facing through it right now as a test, and refine it based on my reaction. Save the template. 5 — The Delegation Test Pick a task I do regularly—I’ll name one if you can’t. Interview me about it until you could execute it without me: inputs, steps, judgment calls, who it goes to, and what failure looks like. Keep asking until YOU are confident, not until I get tired. Then write the runbook, execute the task once with me watching, and incorporate my corrections. Save the runbook. This is the test of everything else we’ve set up. As a bonus, it seems like many people are a fan of the one-shot video note [https://substack.com/@techysurgeon/note/c-271385379?r=3adnor&utm_source=notes-share-action&utm_medium=web] that I wrote recently. Here is a preview of something I’m working on to let people understand the workflow for creating AI videos on their own directly in Claude CoWork: the Higgsfield MCP! Higgsfield is a bit pricey, but if marketing is a part of your business, it seems like it is probably the best tool to rapidly create images and videos, taking advantage of the APIs from the best native tools, including Cadence and GPT’s image generator. Check it out below and let me know if you’d like a deeper dive on how to create AI videos Video below lightly dunking on pathologists was made with ONE PROMPT. Wild. Four Habits to Take With You If you remember nothing else: prompt with CARE—context, audience, role, end product. Don’t accept the first output—it’s the input for the next cycle, and the iteration loop is the whole point. Speak over typing, however clunky it feels at first; you’re not on a timer. And watch other people prompt. I’m a little embarrassed it took me this long to do it deliberately. You’ll learn as much from seeing where someone else gets stuck as from any tutorial—including this one. Your output is going to look different from mine. That’s not a bug. Don’t get frustrated. Just iterate. Like this article on Techy Surgeon? Spread the word! The best part of this newsletter isn’t the tips. It’s the people reading them. Subscribers here have teamed up on projects, swapped workflows, and pushed each other’s work forward. I’ve collaborated with several of you myself. If something here has saved you time or sparked an idea, share it with one person who’d benefit. That’s it! One person. Group subscriptions get a discount. Christian Pean, MD, MS is CEO and Co-Founder of RevelAi Health, Executive Director of AI & IT Innovation at Duke Health, and Assistant Professor of Orthopaedic Surgery at Duke University. He writes the Techy Surgeon newsletter on clinical AI and health policy for surgeons and health system leaders. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit techysurgeon.substack.com/subscribe [https://techysurgeon.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

Claude CoWork Live Artifacts and Scheduled Tasks Tutorial

This is a free preview of a paid episode. To hear more, visit techysurgeon.substack.com [https://techysurgeon.substack.com?utm_medium=podcast&utm_campaign=CTA_7] If you want to go deeper and combine scheduled tasks and artifact with context from meetings. Check out the articles below: How I use AI to create content that can be part of your “AI loops””

The AI Content Flywheel: How I Build an Audience as a Surgeon Without Sounding Like a Robot

This is a free preview of a paid episode. To hear more, visit techysurgeon.substack.com [https://techysurgeon.substack.com?utm_medium=podcast&utm_campaign=CTA_7] Thank you Doug Fullington, MD [https://substack.com/profile/445396101-doug-fullington-md], Alex Rivero [https://substack.com/profile/32407752-alex-rivero], HealthMind Insights [https://substack.com/profile/6308510-healthmind-insights], Eric Burgh, MD [https://substack.com/profile/49261132-eric-burgh-md], and many others for tuning into my live video! Join me for my next live video in the app. I can publish a paper in an upper-tier orthopedic journal and, if I’m being candid, a few hundred to maybe a few thousand people will read it. Some of those people will cite it in their own papers. A small number will change anything in their practice because of it. Then I share an insight about CMS payment model mechanics on Substack, and a health system CFO I’ve never met emails me because she’s been trying to articulate that exact problem to her board. A policy researcher in Geneva follows the thread. Three orthopedic surgeons I’ve never spoken to start a conversation about what I wrote that turns into an ongoing dialogue that has genuinely shaped how I think. That asymmetry in exposure is what motivated me to start writing in public. Not personal branding, not monetization, not the promise of newsletter revenue (though those have their own logic). The primary driver was reach and dialogue. The feeling that the ideas worth communicating in healthcare were reaching a fraction of the people who needed them, mostly because we’d built a science communication infrastructure that made academic publishing the gold standard and everything else secondary. AI changes the equation a bit. Not by doing the thinking. The thinking is still yours, and it’s still the most important part. But by handling enough of the operational needs that someone running a clinical practice, a research agenda, and a startup can also maintain a consistent presence in public. That’s what I want to share here: how I’ve structured that system, in enough detail that you can build your own version of it if it seems useful. What works for me may not work for you. I can’t promise to optimize your content calendar. But I wanted to describe infrastructure that has made the practice of writing in public sustainable for me, and that has produced a lot of unexpected good in the process. Why This Isn’t Primarily About Content Creation A minor but important framing note before the setup details: the value I’ve gotten from Techy Surgeon isn’t primarily the newsletter metrics. It’s the policymakers who’ve reached out. The research collaborators who found me because of a piece on care coordination. The clinicians who wrote to say that an article crystallized something they’d been trying to explain to administrators for years. The people building interesting companies in healthcare who wanted to connect because we were apparently thinking about similar problems from different angles. None of those connections would have happened if I hadn’t been willing to put my thoughts into a form that others could interact with. And it wasn’t a white paper or journal article that did that. It was something closer to a public conversation, where the format invites response and the distribution reaches people outside the academic bubble. I think medicine, and clinical research more broadly, is underinvested in this kind of communication. Not because clinicians don’t have things worth saying (clearly they do), but because the infrastructure for saying them publicly has been either unavailable, considered taboo, or too costly in time. AI is changing that. The skill of generating multimodal media whether written, visual, or video, is becoming something a motivated clinician can build and maintain without a full production team. We probably need more of us doing this. The Flywheel, Briefly Before the setup details, the concept: a content flywheel is a system where each piece of output feeds the next, and where the marginal cost of producing content decreases over time rather than remaining constant. The alternative is what most clinician-writers default to: the brute-force model, where every piece starts from scratch, involves a Sunday afternoon staring at a blank editor, and depends entirely on having energy left over from the clinical and research work. That’s a fragile system. It produces good work occasionally and nothing the rest of the time. The flywheel I’ve built runs on five loops: Ideate (keep a backlog so you never start from nothing), Research (get sources before you start writing, not after), Write (use AI constrained by your voice, not AI in its default state), Distribute (publish once, distribute many times across platforms), and Repurpose (one strong piece seeds two weeks of downstream content). The setup below is how I’ve operationalized each of those loops.

Clinical AI Faceoff: OpenAI's ChatGPT for Clinicians vs OpenEvidence vs DoxGPT

This is a free preview of a paid episode. To hear more, visit techysurgeon.substack.com [https://techysurgeon.substack.com?utm_medium=podcast&utm_campaign=CTA_7] Thank you to everyone who tuned into my live video! Join me for my next live video in the app. I went live at 6:45 the other morning to open three tabs, ChatGPT for Clinicians, Doximity GPT, and OpenEvidence, and ask them the same questions. A few dozen clinicians and subscribers joined at that hour on a Sunday, which I did not expect, and I’m grateful for. The headline finding isn’t who won. It’s that it seems soon you won’t be able to tell the three tools apart from the navigation bar. Each one now has an ambient scribe (or form of one). Each one tracks CME. Each one has a “skills” or “dot flows” tab that, today, mostly amounts to baked prompts dressed up as workflows. OpenEvidence has a feature literally called the dialer — Doximity has had a dialer for a decade. The product surface is converging fast. A quick disclaimer before we go further: opinions here are mine alone. I have no financial relationship with any of these companies. I selected these three because they appear to be getting the most traction in the marketplace — not because they’re the only ones worth your time. Up-to-Date Expert AI, Glass Health, Abridge’s embedded answering, and others all deserve their own look. What each tool is best at right now ChatGPT for Clinicians is the new entrant. Verification is rigorous — NPI, photo of a driver’s license, a ClearID face match — which I read as a deliberate credibility signal. Underneath, the experience is polished but the clinical answers were the weakest of the three on the queries I ran. There is a skills surface that hints at where this is going, but most of the entries today function as prompts rather than true agentic workflows. I did not see a Business Associate Agreement presented during signup, and I have not yet found a satisfying answer on PHI handling. Doximity GPT quietly has the best one-off clinical answers right now. Not by a wide margin, the others are good, but on a hip arthroplasty question and a DVT prophylaxis report, Doximity surfaced the PREVENT CLOT trial [https://www.nejm.org/doi/full/10.1056/NEJMoa2205395] and the CRISTAL trial [https://jamanetwork.com/journals/jama/fullarticle/2802988] at the top of the response, where a domain expert would put them. For a clinician, citation prioritization is trust. Doximity also brings a distribution moat the others can’t replicate quickly — the dialer, fax, telehealth, the news network, and Peer Check (where physician experts grade the answers) — and a redesigned interface that’s the cleanest of the three. OpenEvidence has the lowest friction and the fastest latency. They are clearly throwing serious compute at the answer surface. The differentiator most clinicians never find is Deep Consult. Turn it on, answer two or three follow-up questions, and you get a research-grade brief with embedded figures from JAMA and NEJM, made possible by the licensing partnerships OpenEvidence has signed with NEJM Group [https://www.nejm.org/] and other major publishers. When I asked Deep Consult to brief me on secondary fracture prevention for a quality improvement committee, the output was something I could have walked into a department meeting with that morning. Distribution beats product when the products converge All three are free. All three answer questions credibly (ChatGPT least so). All three are racing to bolt on the same surrounding capabilities. Doximity wins on installed clinician base. OpenEvidence wins on speed, trajectory, raw capability and on Deep Consult. ChatGPT for Clinicians wins, today, on almost nothing — but the verification gate suggests they intend to be taken seriously, and they hold a foundational model and patient facing asset the others don’t. The chat interface is no longer the moat. The moat is whoever first connects grounded clinical evidence to native multimodal output, real workflow extensibility, and physician-earned trust without forcing the clinician to play copy-paste between four tabs to get there. This is the worst these tools will ever be. That should change how we evaluate them: I’m less interested in the question “does it work today?” , and more gravitating to “how do we shape what it becomes?” Date several. Marry none. Use the tool that fits the question in front of you. Send the teams behind them living, breathing feedback. On the Horizon None of these tools is built for patients. The updated guidelines on incidental hepatic steatosis answer ChatGPT gave me this morning was reasonable, and I am a bone surgeon. I read it the way a layperson would, and I would not stake decisions on it without help. The literacy gap between clinical outputs and patient comprehension is not a UX problem. It is a safety problem. Tearing down the gate before we have built tools that respect that gap is how we get harm. The administrative arms race — prior authorization letters, denial appeals, faster note-writing — is a symptom, not a cure. We went deep and fast on the workflows where the money lives, which are the workflows our payer infrastructure forces clinicians to spend their evenings on. That work is valuable, and it is not patient-facing. The places where AI could actually move outcomes — secondary fracture prevention, fall prevention, post-op care navigation, osteoporosis treatment rates that sit around 20% after a fragility fracture [https://www.bonehealthandosteoporosis.org/] when the evidence base for treatment is overwhelming — are still under-resourced. Trust and co-design with clinicians is the unlock. OpenAI has not earned it yet for clinical use. Doximity and OpenEvidence have, in different ways, by being physician-forward from day one. That posture is not optional going forward — it is the moat. The path from clinical intelligence in your pocket to democratized, evidence-based care that actually moves quality and outcomes runs straight through clinicians willing to show up and iterate. That is the dream. We are not there. We can get there. Christian Péan, MD, MHS, is an orthopedic trauma surgeon in Durham, North Carolina. He is core faculty at the Duke-Margolis Institute for Health Policy and CEO and co-founder of RevelAi Health [https://revelaihealth.com/], an AI care management platform for value-based care. Opinions are his own. 🔒 For paid subscribers — the full demo and the operator’s notes The complete screen-share from the morning’s livestream. Side-by-side queries across all three tools, the Deep Consult walkthrough, the prior authorization generation, the acetabular fracture surgical-plan comparison, the live multimodal handoff into a branded HTML committee deck, the connector detour inside Claude, and the clinical trials map I discovered on stage.

Founders Coffee with Techy Surgeon. RevelAi CTO Hadi Javeed, Claude Opus 4.7, and CMS Access

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit techysurgeon.substack.com/subscribe [https://techysurgeon.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]