From one discharge note to a clinic that runs on tools I built.

The first time I used AI for anything real, it wasn’t clinical and it wasn’t clever. I had written a Facebook post about grain-free diets and the DCM heart disease link, and I asked ChatGPT to make it read better. That was the whole ask. I wrote something, make it better.

It did. And then I noticed the same rough notes could become a case summary, a printable, a client explanation. One piece of my thinking, reshaped for whoever needed it.

I didn’t think I had discovered anything. I just stopped rewriting my own posts from scratch.

(The exact date is approximate. Steve, confirm when you want.)

The prompts changed on their own. Instead of “what is this drug,” it became: here’s the patient, here’s the weight, here’s what already happened, here’s what the owner can and can’t do, what would you consider.

I wasn’t engineering perfect prompts. I was dumping the real, messy situation into the conversation and letting it reason with me. Context beat cleverness every time.

That’s when it went from an encyclopedia to something closer to a second brain. Not “diagnose my patient.” More like a colleague who read everything and never gets tired of the question.

Here’s a thing that surprised me. A lot of the value wasn’t translating medicine into plain English. It was translating it between the people in my own clinic.

I make the call. But my receptionist Baylee is the one who talks to the owner. So the prompt became “this is coming from Baylee after she talked to me, say it like that.” Not sounding like she diagnosed the case herself. Not making it sound like we’re begging them to come in.

And my prompting style was never polished. It was “no.” “Shorter.” “Don’t say that.” “They already know that.” “Add gabapentin.” “Okay, put it all together.” Correcting it out loud until it was right is the technique, not a failure of the technique.

At some point I got tired of explaining how Hearthstone runs every single time. Solo doctor, surgery mornings, drop-offs, no overnight staff, the referral options I actually use. Once that context lived somewhere, the prompts got shorter and the answers finally fit how the practice really operates.

Then I hit the wall everyone hits. The AI is smart, and it forgets. A little memory file isn’t the same as knowing your business. I learned, by being annoyed, that intelligence and memory are two different problems.

Memory is infrastructure. You have to build it on purpose. Nobody tells you that until you need it.

Early in 2026 we moved the clinic off our old VIA system onto a modern cloud PIMS with an actual open API. That changed the question. Up to then AI worked on text I pasted in. Now I could ask whether it could touch the real operating system of the practice: appointments, patients, notes, tasks.

So I started experimenting. Reading practice data through the API, seeing what was possible beyond the screens the software gives you. Experiments, not a finished product. Some of it worked, some of it I’m still figuring out.

Around the same time an AI receptionist started fielding calls, with flows I wrote for emergencies, refills, and new clients. The real question stopped being “what prompt do I use” and became “what should AI be trusted to handle on its own, and when does it hand off to a human.”

This is the part I didn’t see coming. For my whole career, if I had an idea for a piece of software, I had an idea. That was it. You find a vendor, you compromise, you pay a lot, or you drop it.

Then the tools got good enough that I could describe a problem, build a rough version, argue with it, break it, fix it, and sometimes have something useful at the end. Not a document. Software that runs.

I’m not a software engineer and I didn’t become one. The threshold just collapsed. The distance between “I wish this existed” and “I made a version of it tonight” got a lot shorter.

This is the one that changed my actual day. I talk through the exam the way I always have. The system listens, drafts the SOAP note and the discharge, and puts them back into the record. I run it on my own hardware.

The hard part was never the audio. It was making the note complete. Early on a discharge-only visit would push the plan and quietly drop the exam findings, the history, the assessment, and still report success. So I rebuilt it to reconstruct the whole note from the transcript, and to tell the difference between me examining a dog and the owner chatting about their weekend.

The demos always look amazing. Tuesday afternoon is where you find out if it actually holds. This one does.

Not everything I build is veterinary, and this one might be my favorite. My dad can hear most conversation, but in a loud cabin with a fan going and dishes clattering he loses the thread. Instead of only shopping for hearing gadgets, I built a small app that shows live captions of what I’m saying, there when he misses something, ignorable when he doesn’t.

I also spent real time on whether calling it “DadCaptions” would make him feel labeled. Building software isn’t just getting the code to run. It’s naming, dignity, and whether the thing makes a person feel small.

Same month I had a bot watching the market and pinging me on Telegram. Turns out the confidence you build in one place carries into the next, and back into the clinic.

2023, I asked an AI to fix a Facebook post. Now I’m talking about APIs, agents, local models, and building software, and none of it came from going back to school. The tools moved toward me.

The next piece is agents with narrow jobs: a clinic-data agent, a records agent, a receptionist agent, each doing one thing well instead of one robot pretending to do everything. That’s what I’m building now, in the open.

That’s what FutureProof.vet is. Not a guru with a master plan. One working vet keeping track of what works, what is ridiculous, and what this means for the rest of us. This is where you come in.