Standing up the Founding Four, for real

Second one.

Last time I ran the early roadmap on one business and it broke in a useful way. This time I ran the whole thing: the full Founding Four, the coordinating layer, the approval gate, on a different real business. It’s a personal-audio retail operation, and I’m keeping the specifics out of this on purpose. The point of this log is the process, not the business. What does it actually feel like to hire three AI cofounders and let them work.

Standing up the roles from the templates.

I didn’t write anything from scratch. I copied the templates straight out of the guide. The shared brief first, what we sell, who we serve, the goals. Then a file per role: Growth, Ops, Brand, each with the sub-files it owns. Then the two coordination files, the chief of staff brief and the approval gate policy. Filling them in is maybe an hour of honest thinking, not a build. That’s the part that still catches me off guard. A role isn’t code. It’s a brief. A role exists the moment its file is real and it has one playbook to run.

The thing that surprised me was writing the approval gate first, before turning anything on. It changed how the whole thing felt. It’s the difference between “an AI is doing stuff in my business” and “a team is drafting, and nothing goes out without me.” Same tools. A completely different amount of trust.

Turning it loose on the first batch.

Then I handed the chief of staff a real batch of work and let it route to the cofounders. Growth drafted outreach and a couple of customer-facing pieces. Ops drafted the weekly read on the business and lined up a few operational tasks. Brand checked all of it against the voice file. One session for me to talk to, three cofounders doing the work behind it.

Here’s what the process actually showed, which is the part worth writing down.

The cofounders drafted. They didn’t act. Every cofounder produced something reviewable and then stopped. Nothing was sent, nothing was posted, nothing was charged. That’s not luck. It’s because the gate says so and the cofounders were briefed to draft-and-hold by default. I opened the session half expecting to catch something overreaching. Instead I spent the whole time reviewing drafts, which is exactly where I wanted to be.

The gate held where it was supposed to. The line I wrote is simple: anything reversible and internal runs on its own, anything that goes live, spends money, or speaks as me stops for a yes. In practice the internal weekly read just appeared, done, while the outreach that would actually land in a customer’s inbox sat in a “waiting on you” pile. I approved some, sent a couple back with a note, killed one. The gate didn’t slow me down. It told me exactly where my attention was actually needed and left me alone everywhere else.

The honesty moment. This is the one I keep thinking about. One draft needed a first-person observation, the kind of “here’s what I noticed this week” line only a human running the business could really say. The cofounder didn’t make one up. It left an explicit [founder] slot in the draft and moved on. It would have been so easy for it to invent a plausible-sounding sentence, and I probably wouldn’t have caught it. It didn’t. That’s the behavior that makes this safe to run: a cofounder that marks the gap instead of papering over it. I’d rather see ten [founder] slots than one confident fabrication.

Where Ops actually lives. The practical discovery. I’d assumed the whole team could run in one place, Cowork, nice and tidy. Not quite. Planning and drafting are happy in Cowork and Chat. But the Ops role’s live actions, the ones that touch real credentials and reach out over the network, want to be in Claude Code, because that’s where the credentials and the network access actually live and where I can watch it work against real files with a diff in front of me. So a split emerged on its own: the team thinks and drafts in Cowork and Chat, and Ops does its hands-on, credential-touching work in Code. I didn’t design that division. The work sorted itself into it.

The takeaway.

The value did not show up when I wrote prompts. It showed up when I reviewed output and cleared approvals. That’s the whole reframe. I spent almost no time crafting clever instructions and almost all of it doing the one job that’s actually mine, reading what the team produced and deciding what goes. A team that drafts everything and holds the risky stuff turns your day into a review queue instead of a prompt-writing session. That’s the thing I couldn’t feel until I ran it.

This is one batch, not ninety days. But the shape is already clear, and it’s the shape the guide promises. Hire the team, brief the gate, then mostly review and approve. More as it runs.