Building my AI operating system changed how I advise healthcare teams

TL;DR

I built my own private AI operating system — with Claude Code and increasingly Codex — despite not having written code in 25 years until recently. The hardest lesson: capture and search are the wrong goals. What earns its keep is a system that assembles what you already know before you need it, without you asking. Stored volume matters far less than timing, and knowing how to shape AI around your actual work is becoming a basic professional survival skill.

I run my AI advisory work on a private operating system I built myself. It lives on my laptop and my phone, I made it with Claude Code and increasingly Codex, and a year ago I had not written code in 25 years. Its core feature is a "second brain", which is why Greg Meyers' post caught my attention this week.

The CDTO of Bristol Myers Squibb wrote that he maintains a personal AI second brain that briefs him before project meetings on what was discussed last time and what is worth probing.

Why capture and search aren't enough

Actually running one of these systems is much less straightforward than it sounds. The instinct is to capture everything. Save the articles, the notes, the theses, the contacts. But a clean searchable pile is still mostly inert if nothing pulls from it when you need it. Capture was never the hard part.

Search does not save you either. It is passive. It only works when you already know what you are looking for, and sitting in front of a client you usually do not. The note that would have sharpened your recommendation is often the one you forgot you saved.

What earns its keep

The version that earns its keep does the opposite of search. Before a meeting, it assembles what I already know about everyone in the room: past interactions, the thesis I held six months ago, the follow-up I missed, the article that suddenly matters again. I do not ask for it manually. It is there when I open the brief. Still imperfect, but already useful enough to have changed how I prepare.

That is the bar the next generation of these tools will be judged on. Stored volume matters much less than timing. Most of what ships next year will still optimise for storage and nicer interfaces. Useful, but not enough.

Building vs. buying

The bigger shift is simpler. I did not buy this system. I built it. That has made me much less patient with AI advice in healthcare that never touches the operating reality of building with it.

Knowing how to shape AI around the work you actually do is becoming a basic professional survival skill. Most people still manage complex work from memory, inboxes and documents. Some adapt themselves to generic tools. The real step change is building systems around how you actually think and work.

Key takeaways

• I built my own private AI operating system using Claude Code and Codex — a year after not having written code in 25 years.
• Capture is never the hard part; a clean searchable pile is still mostly inert if nothing pulls from it when you need it.
• Search is passive — it only works when you already know what you are looking for, and sitting in front of a client you usually do not.
• The version that earns its keep assembles what I already know before a meeting without me asking for it manually.
• Stored volume matters much less than timing — most of what ships next year will still optimise for storage and nicer interfaces, which is useful but not enough.
• I did not buy this system; I built it, and that has made me much less patient with AI advice in healthcare that never touches the operating reality of building with it.
• Knowing how to shape AI around the work you actually do is becoming a basic professional survival skill.