I Set Up Claude Code for 15+ Businesses: The AI Operating System Framework

After the fifth install I stopped improvising and wrote down the shape that kept working. It's four files, in this order, and none of them are optional.

1. IDENTITY — who is this and what do they actually do

Not a mission statement. A plain description of the business, its customers, and the three or four recurring jobs someone on the team does every week. This file is what stops every answer from sounding like generic AI advice.

2. TOOLS — what it's actually allowed to touch

Least privilege, always. A file-organizing agent doesn't need email access. A drafting agent doesn't need to send anything itself. Every install starts with an empty allow-list and I add one tool at a time as the owner asks for it.

3. WORKFLOW — the actual repeatable steps

This is where most setups fail: they describe outcomes ("handle customer emails") instead of steps ("read the inbox, draft a reply, hold it for approval, never send automatically"). Write the steps like you're training a new hire on day one, because that's exactly what's happening.

4. MEMORY — what it's allowed to remember and for how long

A running log of decisions and preferences, reviewed monthly. Without this file every session starts from zero and the owner re-explains the same three preferences forever.

What actually changes between installs

The four-file shape never changes. What changes is scope: a 2-person shop gets one agent covering three jobs, a 20-person team gets three narrower agents that hand off to each other. Start narrow — you can always widen the tool allow-list later, you can't easily walk back an agent that already did something it shouldn't have.

Free resource

Blank IDENTITY / TOOLS / WORKFLOW / MEMORY starter files, ready to fill in for your own setup.

Download the 4-file template