Build Your Own AI Chief of Staff: The 4-Part Starter File

People assume the hard part of building an AI teammate is technical. It isn't. My chief-of-staff agent, Connie, is a plain-language file. No code. I wrote it the way I'd write an onboarding doc for the best assistant I could imagine hiring. And the reason version three runs my day while version one just annoyed me comes down entirely to what's in that file.

So let me give you the structure. Four parts. You can fill it in tonight.

Part 1— Who she is (the role). Not a task, a role. One or two sentences naming what she owns and the standard she holds. Mine starts roughly: "You are my chief of staff. You protect my time, keep my priorities straight, and handle the operational details so I can focus on the work only I can do. You think like a sharp operator who's been with me for years." Notice there's no task in there. There's an identity and a standard. That's what lets her handle the unscripted moments a task list never anticipates.

Part 2— What she knows (the context). This is the part everyone skips, and it's the part that matters most. An AI is only as good as the context it reasons from. So I give her mine: what my week actually looks like, what I'm protecting (my mornings, my podcast hours), how I make decisions, what "good" looks like to me, what I sound like. The first version of Connie was thin here, so she gave thin, generic output and waited on me for everything. The moment I wrote her a real picture of my world, she started making calls I'd have made myself.

Part 3— What she remembers (memory). A teammate who forgets everything every morning isn't a teammate. Give her persistent memory: the ongoing projects, the people, the decisions already made, the patterns. This is what turns "answer this one question" into "carry this forward." It's the difference between an assistant you re-brief daily and one who already knows.

Part 4— When to act on her own (permission). This is the one that changes everything, and the one most people are scared to write. Tell her explicitly when to move without asking: "If you see X, handle it and tell me. Don't wait for me to ask. When it's a judgment call above this line, check with me first." That single section is what flips an agent from "waits for orders" to "takes initiative." It's also, not coincidentally, exactly what psychological safety does for a human team: permission to act without fear.

---

That's the whole file. Role, context, memory, permission.

Here's the honest part, because I promised you the real version. My first draft of this was rough, and so was Connie. I rewrote it three times. Each time I added more real context and more explicit permission, she got more useful. The loop is simple: write it, watch her for a week, adjust the file. Same as managing a person. You don't fire a good hire after one mistake. You give them a clearer guardrail.

And notice what skill this actually is. It's not coding. It's writing down how you work, clearly enough that someone else could step in. You already do this every time you train a new person. The barrier was never technical. It was that nobody told you that the skill is the thing you already have.

If you want the exact file I use, the real one, with the four parts laid out as a template you can drop into any AI chat, get it here

Start with one role. The one you most wish someone would take off your plate. Write it the four parts. Watch it for a week. That's the entire method, and it compounds faster than you'd believe.

-Jonelle