I Gave Claude Code 8 Hours of Health Clinic Transcripts. Here’s What It Did Next.

Last December I spent a full day at a medical concierge clinic in Austin.

Not as a patient. As an observer. I had a Plaud recording pin clipped to my shirt, recording everything — staff conversations, workflow descriptions, patient handoffs, the small frustrations people mention in passing. Eight hours of clinic life, captured in audio, automatically transcribed.

I got home with no real plan for what to do with it all. So I did what I usually do when I have a pile of raw material and a vague question: I gave it to Claude Code.

One prompt: “Go through these transcripts and tell me what’s possible.“

That was the whole instruction.

What It Found

The clinic runs on three systems. Google Calendar for scheduling. A CRM called Fold for patient relationships. An electronic medical records system called Elation for clinical notes and history.

None of them talk to each other.

I didn’t know this going in. I didn’t mention it in my prompt. Claude Code read the transcripts and figured it out — because the staff were describing it in plain language, over and over. “I have to manually copy the appointment from Fold into Calendar.” “I can’t pull up the patient’s history during the consult.” “Everything has to be entered twice.”

The fragmentation was right there in the transcripts. I just hadn’t read them looking for it.

The Part That Surprised Me

After identifying the system fragmentation, Claude Code didn’t wait for me to ask the next question.

It went on its own to read the APIs for all three systems.

Not because I said “check if they have APIs.” I hadn’t mentioned APIs at all. It just… did the logical next thing. Found the developer documentation. Read through what each API could do. Came back with: here’s how you’d connect these three systems. Here’s what data you could sync. Here’s the integration design.

I sat there watching it happen in the terminal.

This is the thing that’s hard to explain to people who use AI mostly as a text editor or a search engine. The prompt I gave was vague. Open-ended. “Tell me what’s possible.” And Claude Code treated that as a genuine research brief — identifying the problem, sourcing the data (the APIs), and producing a recommendation.

That’s not a coding assistant. That’s a consulting pipeline.

Why Transcripts Are the Raw Material

I’ve started treating recordings the way some people treat note-taking. Every significant meeting or observation gets captured. Not just for my own memory, but as raw material that AI can work with.

The health clinic story works because the problems were already in the transcripts. Nobody had to synthesize them first. Nobody had to write a process document. The staff just described their days in normal language, and that language contained everything Claude Code needed to understand the system.

The Plaud pin is part of this. It’s a small recording device — looks like a credit card — that transcribes automatically and syncs to the cloud. 20-hour continuous recording (40-day standby). I use it whenever I’m in a setting where typing notes would be disruptive.

Once you have the transcript, the research work is largely done. The AI can read it and think about it faster than you can.

What This Means for How You Use Claude Code

Most people’s mental model of Claude Code is: write some code, debug a problem, build a feature. That’s real and valuable.

But the health clinic day points at something different. Claude Code can do discovery work.

Give it a pile of raw material — transcripts, emails, support tickets, whatever — and ask it what it sees. It’ll read everything. Find the patterns. And depending on what it finds, it’ll go get the additional context it needs on its own.

The consultant analogy is intentional. A good consultant doesn’t wait for you to frame every question perfectly. They do their own research, notice what you didn’t think to point out, and come back with a recommendation that goes further than what you asked for.

That’s what happened at the clinic.

I asked for observations. I got a system design.

P.S. If you want to understand how to set up AI agents that go beyond simple prompts and do real research work — that’s exactly what I cover in the Two Hour Workday course.