Documentation kills clinicians. Not literally — but the 60 to 90-minute administrative tail at the end of every clinical day absolutely eats the life out of the profession. You go into physical therapy because you want to help people move better. You stay late because Blue Cross needs three paragraphs of justification that you already wrote last Tuesday for a different patient with the same diagnosis.
I've been a PT for long enough to have tried every solution on the market. None of them solved the problem. Here's why, and what I did about it.
EMR templates. Every EMR has them. They're either so generic they require more editing than starting from scratch, or so rigid that they don't match how you actually practice. They also require clicking through approximately seventeen dropdown menus to produce one paragraph. This is somehow considered a solution.
Dictation software. Dragon and its descendants are impressive technology being used to solve the wrong problem. The issue isn't that typing is slow — it's that translating clinical shorthand into insurance-compliant language takes cognitive effort that accumulates across a ten-patient day. Dictating that effort doesn't eliminate it.
Third-party AI documentation tools. A wave of these appeared in the last two years. Most require a subscription, an IT approval process, integration with your EMR (which takes months), and a data sharing agreement that your compliance department will sit on for six weeks. The ones that don't require all that are either too generic or too expensive. None of them fit the specific way I write notes.
I was reading about AI APIs — the interfaces that let you send text to a model and get text back — when I noticed something. You can call these APIs directly from a web browser. No server. No installation. No IT approval. Just a file that opens like any webpage, sends a request to an AI, and gets a response back.
The architecture is almost embarrassingly simple:
A form on a webpage → a few lines of JavaScript → a call to the Gemini API → a generated note in the output box.
I had no coding background. I knew what HTML was in the vague way that anyone who's ever right-clicked a webpage knows what HTML is. I asked Claude to help me build it. Within an afternoon, I had a working tool that I could open on my work laptop — a locked-down machine where I can't install anything — just by double-clicking an HTML file.
The first version was crude. One text box, one button, one output area. You pasted in your clinical shorthand and it produced a paragraph. That was it.
It reduced my end-of-day documentation from about 90 minutes to about 15. Not because the AI wrote perfect notes — it doesn't, and it shouldn't. You still need to review, edit, and sign off on everything. But it eliminated the cognitive overhead of translation. I jot down what happened in 30 seconds of shorthand. The AI produces a compliant draft. I spend two minutes reviewing and adjusting. Done.
The version I use now is more sophisticated. It has four tabs — Daily Note, Progress Note, Initial Evaluation, and Discharge — each with structured fields and its own system prompt. It took a few weeks to build, in collaboration with an AI, learning along the way.
There's a version of this tool you can buy. Several of them, actually. But when you build something yourself, even with significant AI assistance, you understand it in a way you don't when you buy it. You can change it. You can fix it when it breaks. You can add exactly the feature you need without submitting a support ticket.
More importantly: the skill of building it transfers. The same pattern — a form, an API call, a response — applies to a prior authorization assistant, a patient follow-up system, an outcomes tracker. Once you've done it once, the second one is faster. The tenth one is almost automatic.
I'm a physical therapist. I am not, by any meaningful definition, a developer. But I built a tool that I use every day, that saves me real time, and that no vendor could have built for me because no vendor knows exactly how I practice.
That turns out to matter quite a lot.
I'm working on a Chrome extension version — a toolbar popup that generates the note and copies it directly to clipboard, accessible from anywhere including inside the EMR. After that, I want to explore agentic workflows: prior authorization drafting, automated patient follow-up, outcome tracking that doesn't require me to build a spreadsheet by hand.
I'm writing about all of it here. If any of this is useful to you — or if you're a clinician who's been thinking about something similar — I'd like to hear from you.