Infographic by NotebookLMicite your text here...
What's New
May 2026: Publishing your app to Google Cloud became much easier
May 2-26: AI Studio Mobile app launched
June 2026: Creating android apps
June 2026: Design previews and choices
Introduction
Coined in February 2025 by AI researcher Andrej Karpathy, "vibe coding" is a software development paradigm where you use natural language prompts to direct an AI to generate, refine, and debug code. Instead of writing code line-by-line, the human acts as the architect and product manager, while the AI functions as the execution engine. For clinicians, this approach is a game-changer. Clinicians possess the "domain expertise" to know exactly where workflows are inefficient—like a poorly organized chart or a repetitive referral process—but traditionally lack the programming skills to fix them. Vibe coding democratizes software development, allowing doctors to build highly customized solutions tailored to their own administrative pain points and academic tasks.
Below is a table that lists common vibe coding platforms but we will focus on Google's AI Studio. More information can be found in the Resources section.
Table 1: Vibe coding software comparison (courtesy of NotebookLM
Google AI Studio
I will focus on Google AI Studio because it is easy to use, but the reality is that multiple other AI programs can create apps without coding, particularly agentic systems. Examples include Perplexity Labs, Google Canvas, SciSpace, GenSparkAI, Manus AI, etc. Note: AI Studio includes Playground, which highlights many Google programs such as Nano Banana. These programs are more customizable in Playground compared to those offered by Gemini. I will not be discussing Playground, as the third citation in the Resources section addresses both. We will be discussing the "Build" category in AI Studio
Another advantage to using AI Studio is Google is a tech ecosystem because it includes so many applications, such as Google Search, Gmail, Google Earth, Google Drive, Calendar, etc. Google AI Studio allows you to integrate many of these features.
Google AI Studio is intended for prototyping, which means quickly creating a demo to show others or to have them join the effort. It supports text, code, image, video, and even audio prompt inputs, generating advanced outputs such as images, clean JSON, video files, and text responses. With the addition of Antigravity, AI Studio is now a full-stack application, meaning you publish your app and add authentication and other bells and whistles. With a single prompt, it can build the front end (e.g., React), the back end (Node.js), and even provision persistent databases and authentication via Firebase. When you select "publish," Google Cloud hosts your app, and you must also sign up for a Google API key. When you use Gemini 3 Flash Preview, the charges tend to be minimal, as Google encourages people to use it to experiment and build apps. As of June 2026, here are the current model options
Gemini 3.5 Pro (Maximum intelligence): The most capable model for complex reasoning, long-context analysis, and tasks that demand depth over speed. Best reserved for genuinely difficult tasks where accuracy matters most.
Gemini 3.5 Flash (Balanced speed and power): The workhorse model faster than Gemini 2.5 Pro and the default choice for most tasks. Excellent for everyday coding, content generation, and chatbot work.
Gemini 3 Flash Preview (Speed-first lightweight): The fastest and cheapest model. Ideal for simple, high-volume tasks where speed matters more than depth.
It's important to point out two issues. The first is that AI Studio will deploy the app you create on a dedicated website, but it persists only for about 1 week. Access to the app persists, but it no longer has a public URL. This scenario means you will either need to publish to Google Cloud or download the app as a zipped file to GitHub and upload it to Vercel, which has a free hosting (hobby) tier. You can download the app as a zip file by either selecting the gear icon in the upper right, scrolling to the right, and then selecting "GitHub," or by selecting "Code" in the top middle and then choosing the download option. Once you open Vercel, you must identify the GitHub repository you just created. I would recommend that option for most people, as hosting on Google Cloud became easier in April 2026. Furthermore, one of the steps to publish asks if you want to set a monthly financial cap for your app.
It is also important to decide if you need a backend database to store information because that would require AI Studio to add Firebase.
Below is Figure 1, which is the landing page for AI Studio. The steps to build an app are similar to most prompts. 1. What is the input? A Word document or PDF. A web link? 2. What computation is needed by your app? 3. What is the output? A CSV file? A text file, Word document, or PDF file?
The process is very iterative. Begin with the basics as described above and continue tweaking by asking Gemini questions. As you progress, you will see that Gemini makes recommendations on its own, such as "Would you like me to add a reset button?" There is a section called "Exercise" where we build an app for better explanation.
Figure 1: Home page for Google AI Studio
Clinical Apps
While vibe coding makes prototyping easy, clinical applications require strict adherence to HIPAA and other regulations.
1. AI Studio is for Prototyping (No PHI) Standard public configurations of Google AI Studio and free-tier accounts are not HIPAA-compliant and must never process Protected Health Information (PHI). Data entered into the free tier may be subject to human review and used to train future public models. Therefore, AI Studio should be used as an "innovation funnel"—a place to quickly build and validate Minimum Viable Products (MVPs) using de-identified or dummy clinical data.
2. Migrating to Enterprise Environments for Production If a prototype proves useful and needs to handle real patient data, the pipeline must be rebuilt or migrated to an enterprise-grade environment like Google Cloud’s Vertex AI. Vertex AI provides a Business Associate Agreement (BAA), ensures zero-data retention for model training, and supports Customer-Managed Encryption Keys (CMEK) and VPC Service Controls to prevent data exfiltration. Professional developers often warn that "vibe coding" your way straight into a production healthcare app is a massive liability without these hardened security infrastructures.
3. The 3-Tier Risk Stratification Framework
Tier 1 (Personal Tools): Calculators or personal workflow apps with no patient data (e.g., the injection calculator). The clinician can test and validate the app themselves.
Tier 2 (Administrative/Team Tools): Department scheduling or standardized care plans. Requires review by another clinician and notification to IT governance.
Tier 3 (Patient-Facing/Decision Support): Tools that directly influence clinical decisions or handle PHI. These require formal institutional review, professional software developer input and may qualify as regulated medical devices.
Key Takeaways for Clinicians
1. "All-in-One" Browser Builders (Best for Non-Technical Clinicians) Tools like Google AI Studio, Replit, and Lovable bundle the AI, the code editor, and the hosting environment into a single browser window. Clinicians can describe what they want, see a live preview, and launch the app to a live URL without ever installing software on their computer.
2. AI-Assisted IDEs (Best for Clinicians with Coding Experience) Tools like Cursor, Claude Code, or Google's Gemini Code Assist act as autonomous "pair programmers." They sit inside a traditional coding interface and help write, debug, and execute commands across multiple files. While powerful, they still require the user to understand software architecture, terminal commands, and how to host an application.
The "Security Debt" Warning for Healthcare Regardless of the platform chosen, you should emphasize that vibe coding often skips traditional security reviews and code validation. Relying purely on AI can introduce "security debt"—such as hardcoded passwords or data exposure vulnerabilities (as seen in some Lovable apps). Clinicians should treat vibe-coded apps as prototypes only and ensure they are properly vetted and migrated to hardened, HIPAA-compliant enterprise environments before ever processing real patient data
Academic Apps
Once you start creating academic apps, you will begin to appreciate the need for more apps, the ease at which they can be built, and how one good idea begets another. The following are five examples of academic apps that were built using Google AI Studio.
We are evaluating a technique to clone or augment existing small datasets. While we have evaluated various new AI techniques to augment data, we found that none were adequate, largely due to the lack of statistical similarity. We used AI to research potential methods and discovered an older technique that is rarely discussed. The technique uses bootstrapping and adaptive Gaussian noise to create data that is statistically very similar but has no duplicates. I have named this technique BootsMD. Gemini successfully created an app that utilized this technique, which included a necessary upload window, an option to select the size of the augmented dataset (1000-5000), and an automatic download feature for the dataset.
I then thought, why not add the statistical tests at the same time to show the degree of similarity? Now, when the data is generated, so is a report that provides descriptive statistics, visualizations, and various statistical tests for numerical and categorical data. You can download the augmented data as a CSV file and the statistical report as a text file.
A biostatistician is in the process of validating our results using the statistical package SAS. Our goal is to publish our work in the near future.
Figure 2: The input and size options
Figure 3: The outputs
It is rare to find a graph database (knowledge graph) (KG) used in medicine, as it is one of the last industries to embrace this new way of storing data instead of using a relational database. Many others, such as Netflix and Walmart, leverage knowledge graphs. In addition, it is really rare to find a demo or live knowledge graph (KG) that you can mine (query). Many people need to learn a new programming language, such as Cipher, to do so. Because medical knowledge graphs are rare, there are no demos to show students, and therefore the project was an opportunity to build one.
First, I asked Gemini if it could create a knowledge graph (KG), and it answered in the affirmative, so I asked it to create a small medical KG. It did that successfully, so I then asked it to use real medical data. On its own it said, "I will use data from the Cleveland Clinic Heart Study." It added the filters you see on the left, and it added shading to suggest those with no heart disease belonged in the greenish area and those with heart disease belonged in the pink zone. It also clusters the data by age. It also added an "Ask Gemini" option to query the database and an export CSV button to download answers. Lastly, it created visualization and correlation buttons at the top of the app. Each blue circle is a node, and that represents a patient. When you mouse over a node, you can see patient characteristics.
Figure 4: Medical knowledge graph
Almost every current large language model (LLM) will perform a critical review of scientific articles. What's different about this one is that it was given additional instructions to screen for evidence that this paper was generated by a paper mill.
An app like this one could also be designed to follow specific journal guidelines (e.g., Elsevier) as well as international guidelines, for example, TRIPOD-LLM. In the lower left image, you can see the graphical user interface where you upload an article for analysis; in the lower right, the results show a very low score (0/10), indicating this paper was likely not written by a paper mill. This app could be combined with the one I created to screen references. There are lots of other additions that could be made, such as author verification, etc.
Figure 6: Results and score
Figure 5: Upload screenshot
The references in scientific articles can be incorrect by accident or hallucinated by an LLM. Importantly, they can be misleading or irrelevant when generated by a paper mill. I wanted to create a way to screen references, primarily for accuracy. Gemini recommended that I verify references by accessing the following resources: DOI, PubMed, Scopus, Web of Science, and the Directory of Open Access Journals. That worked fine but missed pre-print journals. I added the preprint journal arXiv, which is particularly important for computer science-related/data science/AI-related articles, and I made a few minor tweaks. The end result is that it checks these different resources before rating the references as low, medium, or high risk. In the lower left is the page where you upload the article, and the report appears in the lower right. The report is downloadable.
Figure 8: Analysis report
Figure 7: Upload PDF
There are several reasons why you might want to generate synthetic patients. First, they could be used for teaching how to conduct a proper history and physical exam. Secondly, they could be used to teach medical coding. Third, synthetic patients can be used to train machine learning and AI models, altough the number of synthetic patients rquired would be high.
Figure 9 shows the standard scenario choices for a patient. Figure 10 displays the results of a patient with acute pancreatitis due to alcohol abuse. Note the tabs for history, exam, labs, and diagnosis/plan. The synthetic patient case can be downloaded. Other features could be added so as ICD and CPT codes.
Figure 10: Case report with tabs for different sections
Figure 9: Input parameters
Build Your First App
We will be building an automated patient education handout generator. Here are the steps to get started:
Select Gemini 3 Fast Preview as your LLM
Prompt "You are an expert clinical communicator. Translate complex medical diagnoses into empathetic, clear, jargon-free patient education handouts based on the requested reading level. Include sections for: What it is, Common symptoms, Next steps, and When to call the doctor. Create a window for diagnosis and another window for educational level."
Note: AI Studio often includes useful text to the right of your prompt. If you like their addition use the tab key to add
Select the build button and wait
Look at "design preview" and decide whichdesign looks best
Examine the final product and make changes
Understand that errors may occur, but Gemini does excellent problem-solving. Copy and paste the error message back into the prompt.
Also, you might build an app when AI Studio is super busy, and they may ask you to try later. Such an event occurs rarely.
The patient handout generator I created included tone of voice. Be sure it allows you print out the handout
Did you request the option to email a handout to patients using Gmail?
What other options would be important to you?
Resources
Nineteen vibe coding apps: https://www.cio.com/article/4165921/19-vibe-coding-tools-for-democratizing-app-development.html
Official Google Documentation for AI Studio https://ai.google.dev/gemini-api/docs/aistudio-build-mode
Quick Start: Basic instructions
Build Mode & Vibe Coding Guide
Developing Full-Stack Apps
Building Android apps
Deploying your app
FAQs
AI Studio: Playground and Build Functions https://youtu.be/QDLAhBVF_6g?si=XizkhZxtTSbg3c1a
Data Camp. Google AI Studio Tutorial: Complete Guide to Chat, Build, and Stream Modes. https://www.datacamp.com/tutorial/google-ai-studio-tutorial