what changed for clinicians of any background building their own software with ai — coding agents, models, health data access, privacy & the rules — checked daily, each note against at least two sources
You do not need to be a programmer to build a useful tool for your own work. You need a coding agent, a small problem you understand well, made-up data to practise on, and a clear idea of where the lines are on patient privacy and regulated software. These five steps get you from nothing to a first working tool, safely. None of it is quick or effortless, but none of it needs a computer science degree.
Step 1
A coding agent is a program you talk to in plain language. It reads the files in a folder, writes code, runs it, sees what broke and fixes it, while you say what you want and check the result. Some run in a terminal (the text window where you type commands), some inside a code editor, some in the browser.
Pick one, install it following its own instructions, and ask it to explain each step as it goes. Telling it you are new to coding is useful information, not an admission. The Coding agents layer below tracks what has changed in these tools and what they cost.
Step 2
Start with a problem from your own week that is annoying, repetitive and low-stakes: a calculator you look up every shift, a checklist, a way to tidy a spreadsheet you already export, a patient handout generator. Describe it to the agent in a few sentences and ask for the simplest version that works on your own computer.
Use invented data only. Synthea generates realistic synthetic patient records, so you can test on something shaped like real charts without touching real ones. Keep each project in its own folder, and ask the agent to set up version control (git) so every change can be undone.
Step 3
In the US, if you work for a covered entity (most practices, hospitals, health plans), patient information you handle is protected health information under HIPAA. Sending it to an outside service, including an AI model's API, generally requires that service to have signed a business associate agreement (BAA) with your organization. A consumer chat account or a personal API key usually does not come with one.
Until you have confirmed what applies to you, with your organization's privacy or compliance office, keep real patient data out of every prompt, file and tool you build. Models that run entirely on your own computer avoid sending data out, but they do not remove your other HIPAA duties. Also keep secrets such as API keys out of your project files: coding tools and their add-ons have been a target, as several notes here show.
Step 4
Some software that supports clinical decisions is regulated by the FDA as a medical device; some is not. FDA's clinical decision support guidance describes where that line falls, and a few states now add their own rules for AI used in care. Whether a tool you built crosses the line depends on what it does and how it is used, not on who wrote it.
Tools for your own administrative work, education or reference are a different matter from tools that analyse a specific patient's data to recommend treatment. If yours is heading toward the second kind, or you plan to share it beyond yourself, read the guidance and talk to your organization's regulatory or legal team before you deploy it.
Step 5
Once a tool works on fake data, the next step is usually health records. FHIR is the standard format most EHRs use to share data, and SMART on FHIR is the way an app signs in to an EHR with a user's permission. Both have public test environments, so you can practise connecting without touching a live system. Epic, the largest EHR vendor, runs its own developer site.
Things change fast in every layer here. The feed on this site is rechecked daily; the stack pages say where each layer stands now.
Links on this page last checked 2026-11-22.