Building an "AI operating system" for a small team — and I keep hitting the same architecture question. Curious how others have solved it.

The setup: small office (think 3–4 people, mixed technical ability). One person (lets say me but its for client ) has the power-user seat — Claude Code, all the skills, the full context. The others need to feed information in and benefit from the skills, but they're not going to live in a terminal.

Where I've landed so far — call it "one brain, many feeders":

• A shared brain (Drive folder) holds context, intake, and output

• I sit in "mission control" with the one Claude Code seat and run the skills

• Everyone else just drops structured files into intake folders — no terminal, no setup

• Claude validates, organizes, drafts, flags what's missing — a human does the final publish/send

Two principles that keep proving themselves:

1. Force the data + its context to arrive together (one folder per item, a fill-in template beside each file). Orphaned data is what kills these pipelines.

2. Automate the 95%, keep a human on the last 5%.

What I'm chewing on:

• Is per-person seats actually better than one operator + feeders? When does it flip?

• Where should the shared brain live — Drive, repo, DB, something else?

• How do you get non-technical teammates to feed it *consistently*?

If you've built something like this — what's your stack and what did you learn the hard way? Genuinely want to hear implementations, not just theory.

7 comments