Vibecoding is likely the most important skill of our lifetime. It is literally the ability to talk to computers in English. This is a guide for physicians, by a physician who loves to vibecode.

"Vibecoding" is the art of building software by describing what you want (the vibe, the outcome, the function) rather than writing the how (syntax, semicolons, memory management).

The goal of this guide is to introduce you not just to the tools, but to the mindset and framework. That is 90% of the work and where the magic lies. The tools will keep improving, but they remain only as strong as the thinking you bring to them.

This is practical, not theoretical. I have personally vibecoded over 150 apps, many of which I currently use in my own daily clinical practice. This guide is the distillation of that experience—taking the tools of software engineering and translating them into a workflow that fits the physician's mind.

As a physician, you are already a vibecoder. You write an order: "Give 1g Ceftriaxone IV daily." You don't manufacture the molecule or cannulate the vein yourself; you provide the intent and the parameters. AI coding tools are your nurses and pharmacists—they handle the execution.

Before you open the toolbox, you need to adopt the right mindset.

1. Communicate Clearly:

   • Physicians often process thousands of data points subconsciously to reach a conclusion. We are trained to be efficient with our words. Context is key.

   • The Shift: LLMs cannot read your intuition. You must explicitly state what you want to achieve and the why behind it. Be verbose. Be specific. Don't just say "fix it"; explain the clinical context.

2. Think in Systems:

   • Don't just look at the code snippet in isolation. Think about how each component connects, similar to how you would map a disease process to its source and then build a comprehensive treatment plan.

   • Why this matters: Systems thinking prevents bloat and reduces extra clicks. It helps you decide when a basic calculator logic works better than a massive LLM, or conversely, where an AI tool can augment a simple text field to provide significantly more functionality and value.

   • The Shift: Ask yourself: "If I change this input, how does it affect the database? Is this the most efficient tool for this specific task?"

3. Iterate, Iterate, Iterate:

   • The first response is rarely the final diagnosis. It can be frustrating at first, but you must push the LLM further.

   • The Shift: If the AI fails, push back. Ask secondary and tertiary questions. By refining your prompt, you aren't just fixing the code; you are learning how to speak the language of AI. Trust me, once you get the hang of this, it becomes a superpower you will use in every interaction with AI.

Think of your AI toolkit not as a single computer program, but as a stationary box (yes I am old school and I love stationary boxes, still use a few regularly). You wouldn't use a highlighter to sign a prescription, and you wouldn't use a geometry set to write a quick note.

Best for: Brainstorming ideas, asking "what if," and rapid prototyping that you might throw away. They are also great for writing system prompts for any LLM or agent integrated into your app.

These are the chat interfaces. They are low-stakes environments. You can scribble an idea ("Make a calculator for creatinine clearance"), see if it works, and erase it (start a new chat) if it doesn't.

• ChatGPT (OpenAI): The standard. Reliable, good for general logic and brainstorming.

• Claude Chat (Anthropic): A sharper pencil. Excellent at nuanced medical instructions and "thinking" through a clinical pathway before writing code. Found it to excel at complex logic such as shift schedule management.

• Gemini (Google): A pencil with a built-in real time encyclopedia. Great for uses that require up-to-date information or huge amounts of context (like pasting in a whole PDF guideline).

Best for: Building the actual app that you will send to a colleague or use in the clinic.

• Replit: Start Here (The Complete IDE).

  • Best for: Learning, prototyping, and launching real apps without headache.

  • Why: It is a complete Integrated Development Environment (IDE) that lives in your browser. It handles the "hard parts" that usually stop beginners:

    • The Agent: It features a comprehensive Replit Agent that goes beyond just writing code. It has web search capabilities, meaning you can ask it to "Find the latest formula for creatinine clearance and build a calculator for it," and it will research the context before coding. It acts as a project manager, researcher, and coder in one.

    • Security: It manages vital security configurations and environment variables automatically, crucial for those new to coding who might otherwise leave keys exposed.

    • Database: It solves the trickiest barrier to entry: setting up and connecting a database happens instantly in the background.

    • Deployment: One click to go live. No server management required.

  • My Experience: I've made over 130 apps on Replit and they keep shipping new improvements and features such as the recent ability to publish an iOS app directly to the app store. I have also tried nearly all of the other popular builders, most are cheaper, but found them lacking in the infrastructure support Replit gives. Especially as a physician lacking time to spend learning each element.

• Claude Code / CoWork / OpenCode / Codex:

Best for: Finding specific information, debugging, and understanding the "Why."

• ChatGPT Deep Research / Google Deep Research: The heavy-duty researchers. Unlike a standard Google search, these tools will read through multiple sources, aggregate data, and verify facts. Use them to ensure your app's medical logic is based on the absolute latest trials or guidelines.

• NotebookLM (Google): The "Smart Binder."

  • Best for: Digesting your own library of PDFs, guidelines, and lecture notes.

  • Why: Unlike web search, this answers questions only based on the documents you upload. It prevents hallucinations by grounding the AI in your specific textbook or hospital protocol. It can even turn your papers into a listenable "podcast" for your commute, reports and also slide decks and infographics.

Best for: Drafting blueprints, complex solutions and system prompts.

• Google AI Studio: (Also lives here). Excellent for testing how an AI responds to specific medical scenarios before you build the full app.

• ChatGPT 5.2 Pro: The "Reasoning Engine."

  • Best for: Verifying complex clinical logic, massive architecture planning, and long-context analysis.

  • Why: With its "xhigh" reasoning effort and 400k context window, it excels at checking your "math" and ensuring structural integrity before you build. It acts as the master architect.

Best for: Making it look good (Frontend/UI) and creating assets in minutes.

• Figma / Framer: These are pure design tools. Before you ask the AI to build the house, you use these to draw the floor plan. They help you visualize the "Look and Feel" (UI/UX). Use their built in capabilities to connect with Claude Code or Replit or any other IDE.

• Nano Banana Pro (Gemini 3 Pro Image): The "Digital Illustrator."

  • Best for: Generating high-fidelity medical diagrams, patient education pamphlets, or marketing assets. Unlike older models, it handles text generation within images perfectly (e.g., correctly labeling a diagram of the heart).

• Veo 3.1: The "Videographer."

  • Best for: Creating short, high-quality video clips for patient education (e.g., "Show an animation of how a knee replacement works"). It allows you to generate custom video assets on demand to embed in your apps.

One frustrating but important area to cover. Before you build, you must understand the rules of the road. I truly believe that for healthcare to really make the most of AI as a technology, HIPAA regs need to be innovated on, but thats a discussion for another day.

1. The Red Zone (NOT Compliant):

   • Standard ChatGPT / Claude / Gemini / Replit (Personal Plans).

   • Rule: Never paste real patient names, DOBs, or MRNs here. Use "John Doe" or synthetic data only.

2. The Green Zone (Can Be Compliant):

   • Local LLMs: Running an AI (like Llama 3/Qwen) entirely on your own laptop using OpenCode/Codex. The data never touches the internet.

   • Enterprise Cloud: Google Cloud (Vertex AI) or Azure OpenAI IF your organization has a signed BAA (Business Associate Agreement).

3. A Disclaimer:

   • Just because you can build it, doesn't mean it's legal to use. Always consult your hospital's IT/Compliance officer before deploying any tool that touches PHI (Protected Health Information).

Note: Use dummy data (e.g., "Patient X") for this workflow.

This workflow shows how you use the entire Stationary Box to build one cohesive tool.

1. Sketch & Research (The Prep Work):

   • Concept (Pencil): Open ChatGPT, Claude, or Gemini.

     • Prompt: "I want to build a meds reminder for elderly patients. What are the key features needed for accessibility? Help me write a robust system prompt for a coding agent."

   • Data (Highlighter): Use Deep Research.

     • Prompt: "Find the exact dosage guidelines and common side effects for the top 10 most common statins. I need this data to populate my app's database."

   • Assets (Markers): Create your multimedia.

     • Images: Use Nano Banana Pro to create a friendly, high-fidelity icon of a pill bottle with the app name on it.

     • Video: Use Veo 3.1 to generate a 5-second loop of a 'confetti celebration' to play when patients finish their meds.

2. Design Mode (Visuals First):

   • Action: Move to Replit Agent.

   • Prompt:

Vibecoding is just rounding on your software. Going through your census and iterating based on each biomarker or page that comes in.

There is much much more...MCPs, agents, skills, subagents, Clawdbot. In time. Will be coming out with more content and tools on this in coming weeks. If you have ideas or want specific tools covered, feel free to reach out.

Also, I have no affiliation with any of the tools above, nor paid by them.

Vibecoding is likely the most important skill of our lifetime. It is literally the ability to talk to computers in English. This is a guide for physicians, by a physician who loves to vibecode.

"Vibecoding" is the art of building software by describing what you want (the vibe, the outcome, the function) rather than writing the how (syntax, semicolons, memory management).

The goal of this guide is to introduce you not just to the tools, but to the mindset and framework. That is 90% of the work and where the magic lies. The tools will keep improving, but they remain only as strong as the thinking you bring to them.

This is practical, not theoretical. I have personally vibecoded over 150 apps, many of which I currently use in my own daily clinical practice. This guide is the distillation of that experience—taking the tools of software engineering and translating them into a workflow that fits the physician's mind.

As a physician, you are already a vibecoder. You write an order: "Give 1g Ceftriaxone IV daily." You don't manufacture the molecule or cannulate the vein yourself; you provide the intent and the parameters. AI coding tools are your nurses and pharmacists—they handle the execution.

Before you open the toolbox, you need to adopt the right mindset.

1. Communicate Clearly:

   • Physicians often process thousands of data points subconsciously to reach a conclusion. We are trained to be efficient with our words. Context is key.

   • The Shift: LLMs cannot read your intuition. You must explicitly state what you want to achieve and the why behind it. Be verbose. Be specific. Don't just say "fix it"; explain the clinical context.

2. Think in Systems:

   • Don't just look at the code snippet in isolation. Think about how each component connects, similar to how you would map a disease process to its source and then build a comprehensive treatment plan.

   • Why this matters: Systems thinking prevents bloat and reduces extra clicks. It helps you decide when a basic calculator logic works better than a massive LLM, or conversely, where an AI tool can augment a simple text field to provide significantly more functionality and value.

   • The Shift: Ask yourself: "If I change this input, how does it affect the database? Is this the most efficient tool for this specific task?"

3. Iterate, Iterate, Iterate:

   • The first response is rarely the final diagnosis. It can be frustrating at first, but you must push the LLM further.

   • The Shift: If the AI fails, push back. Ask secondary and tertiary questions. By refining your prompt, you aren't just fixing the code; you are learning how to speak the language of AI. Trust me, once you get the hang of this, it becomes a superpower you will use in every interaction with AI.

Think of your AI toolkit not as a single computer program, but as a stationary box (yes I am old school and I love stationary boxes, still use a few regularly). You wouldn't use a highlighter to sign a prescription, and you wouldn't use a geometry set to write a quick note.

Best for: Brainstorming ideas, asking "what if," and rapid prototyping that you might throw away. They are also great for writing system prompts for any LLM or agent integrated into your app.

These are the chat interfaces. They are low-stakes environments. You can scribble an idea ("Make a calculator for creatinine clearance"), see if it works, and erase it (start a new chat) if it doesn't.

• ChatGPT (OpenAI): The standard. Reliable, good for general logic and brainstorming.

• Claude Chat (Anthropic): A sharper pencil. Excellent at nuanced medical instructions and "thinking" through a clinical pathway before writing code. Found it to excel at complex logic such as shift schedule management.

• Gemini (Google): A pencil with a built-in real time encyclopedia. Great for uses that require up-to-date information or huge amounts of context (like pasting in a whole PDF guideline).

Best for: Building the actual app that you will send to a colleague or use in the clinic.

• Replit: Start Here (The Complete IDE).

  • Best for: Learning, prototyping, and launching real apps without headache.

  • Why: It is a complete Integrated Development Environment (IDE) that lives in your browser. It handles the "hard parts" that usually stop beginners:

    • The Agent: It features a comprehensive Replit Agent that goes beyond just writing code. It has web search capabilities, meaning you can ask it to "Find the latest formula for creatinine clearance and build a calculator for it," and it will research the context before coding. It acts as a project manager, researcher, and coder in one.

    • Security: It manages vital security configurations and environment variables automatically, crucial for those new to coding who might otherwise leave keys exposed.

    • Database: It solves the trickiest barrier to entry: setting up and connecting a database happens instantly in the background.

    • Deployment: One click to go live. No server management required.

  • My Experience: I've made over 130 apps on Replit and they keep shipping new improvements and features such as the recent ability to publish an iOS app directly to the app store. I have also tried nearly all of the other popular builders, most are cheaper, but found them lacking in the infrastructure support Replit gives. Especially as a physician lacking time to spend learning each element.

• Claude Code / CoWork / OpenCode / Codex:

Best for: Finding specific information, debugging, and understanding the "Why."

• ChatGPT Deep Research / Google Deep Research: The heavy-duty researchers. Unlike a standard Google search, these tools will read through multiple sources, aggregate data, and verify facts. Use them to ensure your app's medical logic is based on the absolute latest trials or guidelines.

• NotebookLM (Google): The "Smart Binder."

  • Best for: Digesting your own library of PDFs, guidelines, and lecture notes.

  • Why: Unlike web search, this answers questions only based on the documents you upload. It prevents hallucinations by grounding the AI in your specific textbook or hospital protocol. It can even turn your papers into a listenable "podcast" for your commute, reports and also slide decks and infographics.

Best for: Drafting blueprints, complex solutions and system prompts.

• Google AI Studio: (Also lives here). Excellent for testing how an AI responds to specific medical scenarios before you build the full app.

• ChatGPT 5.2 Pro: The "Reasoning Engine."

  • Best for: Verifying complex clinical logic, massive architecture planning, and long-context analysis.

  • Why: With its "xhigh" reasoning effort and 400k context window, it excels at checking your "math" and ensuring structural integrity before you build. It acts as the master architect.

Best for: Making it look good (Frontend/UI) and creating assets in minutes.

• Figma / Framer: These are pure design tools. Before you ask the AI to build the house, you use these to draw the floor plan. They help you visualize the "Look and Feel" (UI/UX). Use their built in capabilities to connect with Claude Code or Replit or any other IDE.

• Nano Banana Pro (Gemini 3 Pro Image): The "Digital Illustrator."

  • Best for: Generating high-fidelity medical diagrams, patient education pamphlets, or marketing assets. Unlike older models, it handles text generation within images perfectly (e.g., correctly labeling a diagram of the heart).

• Veo 3.1: The "Videographer."

  • Best for: Creating short, high-quality video clips for patient education (e.g., "Show an animation of how a knee replacement works"). It allows you to generate custom video assets on demand to embed in your apps.

One frustrating but important area to cover. Before you build, you must understand the rules of the road. I truly believe that for healthcare to really make the most of AI as a technology, HIPAA regs need to be innovated on, but thats a discussion for another day.

1. The Red Zone (NOT Compliant):

   • Standard ChatGPT / Claude / Gemini / Replit (Personal Plans).

   • Rule: Never paste real patient names, DOBs, or MRNs here. Use "John Doe" or synthetic data only.

2. The Green Zone (Can Be Compliant):

   • Local LLMs: Running an AI (like Llama 3/Qwen) entirely on your own laptop using OpenCode/Codex. The data never touches the internet.

   • Enterprise Cloud: Google Cloud (Vertex AI) or Azure OpenAI IF your organization has a signed BAA (Business Associate Agreement).

3. A Disclaimer:

   • Just because you can build it, doesn't mean it's legal to use. Always consult your hospital's IT/Compliance officer before deploying any tool that touches PHI (Protected Health Information).

Note: Use dummy data (e.g., "Patient X") for this workflow.

This workflow shows how you use the entire Stationary Box to build one cohesive tool.

1. Sketch & Research (The Prep Work):

   • Concept (Pencil): Open ChatGPT, Claude, or Gemini.

     • Prompt: "I want to build a meds reminder for elderly patients. What are the key features needed for accessibility? Help me write a robust system prompt for a coding agent."

   • Data (Highlighter): Use Deep Research.

     • Prompt: "Find the exact dosage guidelines and common side effects for the top 10 most common statins. I need this data to populate my app's database."

   • Assets (Markers): Create your multimedia.

     • Images: Use Nano Banana Pro to create a friendly, high-fidelity icon of a pill bottle with the app name on it.

     • Video: Use Veo 3.1 to generate a 5-second loop of a 'confetti celebration' to play when patients finish their meds.

2. Design Mode (Visuals First):

   • Action: Move to Replit Agent.

   • Prompt:

Vibecoding is just rounding on your software. Going through your census and iterating based on each biomarker or page that comes in.

There is much much more...MCPs, agents, skills, subagents, Clawdbot. In time. Will be coming out with more content and tools on this in coming weeks. If you have ideas or want specific tools covered, feel free to reach out.

Also, I have no affiliation with any of the tools above, nor paid by them.