How to build an AI chief of staff, step by step.
To build an AI chief of staff, give one AI assistant standing context about you, connect it to your email and calendar so it can read them, have it brief you every morning, then teach it one recurring task at a time. Add rules for what it may never do, put it on a schedule, and add more agents only when one assistant is doing too many jobs.
I run my own AI chief of staff: 17 agents and one board. I started with one assistant and added agents as I needed them. Steps 1 to 8 are the free 30-Day AI Starter Plan, with no code. Steps 9 and 10 are how I grew it into a team.
-
1. Map where your time actually goes
Before you build anything, write down everything that asks for your attention in a normal week and how long each one takes. Point the assistant at the biggest one first.
For me it was my inbox. Things were landing in email, texts and three calendars, and I was the one keeping track of all of it. So I mapped where the time actually went, cut what didn't need to happen at all, and built for what was left. It's the same Lean order I use with clients: fix the process first, then point AI at it.
-
2. Give it standing context
Create one workspace with standing instructions (a Project in Claude or ChatGPT) and fill it with a one-page brief: who you are, what the business does, who your customers are, how you like things written, and what it should never do without asking.
This is the single biggest upgrade available to you, because you stop re-introducing yourself in every conversation. Name the workspace "Chief of Staff." That name is a decision about what you're building. The context interview prompt writes the brief with you, one question at a time. Projects are included on Claude's free plan.
-
3. Connect your email and calendar, read first
Turn on the connectors for your email and calendar, and read every permission screen. You are deciding what it can see, and you can disconnect any time.
Start with reading, not sending. Ask "What's on my calendar this week?" and "What's sitting in my inbox that actually needs me?" and notice that you pasted nothing. Connectors are on Claude's free plan, so you can do this without paying anything.
-
4. Run a morning brief every day
Have it read your calendar and inbox each morning and write one page: the three things that matter most today, what needs you before noon, what can slip, and one thing you appear to be forgetting.
Read it with your coffee for a week. This is the habit everything else builds on, and it shows you what context is still missing. Here's the morning brief prompt.
-
5. Automate one recurring task
Pick the recurring task you complained about most recently, do it with the assistant once, then turn what you did into a reusable recipe. One task, not five.
Frequency beats size, so a ten-minute task you do every day is worth more than a two-hour one you do once a month. Time it before and after and write both numbers down. Most people who finish the plan find 3 to 5 hours a week.
-
6. Close out each day
Spend the last two minutes of the workday telling it what happened, in fragments, and have it update one running to-do list: what's done, what's new, and anything you promised someone.
By midweek the morning brief opens with what yesterday's close left unfinished. Keep one list as the source of truth and have everything else read from it, so you never have two versions of what's still to do.
-
7. Write the rules of the road
Write down what it drafts but never sends, what information never goes in, and how you'll spot-check its work. Nothing leaves without your eyes on it: AI drafts, you send.
Some things never go in at all: customer personal details, financial credentials, anything a lawyer is involved in. And trust is earned per task, not granted to the system. Let a workflow prove itself five times before you stop checking every line, and keep checking now and then, forever. The guardrails guide goes deeper.
-
8. Put it on a schedule
Turn the morning brief into a scheduled task that runs every weekday before you start, so it's ready when you sit down. Scheduled tasks are on Claude's paid Pro plan.
Scheduled runs need stricter rules than the ones you run yourself, because nobody is there to answer a question. Mine never ask, never send anything, and say exactly what failed when something does.
-
9. Add agents when one assistant is doing too many jobs
When one set of instructions starts covering too many jobs, split it: one agent per job, each with its own instructions, and one supervisor agent that decides which agent a request belongs to.
The rule I use is one owner per job. Only one agent is allowed to clean up email, only one keeps the running list, and every other agent reads from them instead of keeping its own copy. I have 17 agents now, across my calendar, my ventures and my house, and they all report to one supervisor. Here's how those 17 are organized.
-
10. Give it a board and a log, then check it every week
Once several agents are working, add a shared change feed they write to, a one-line log for every scheduled run, and one board that shows you what's still to do and what's done.
Agents need to know what the others changed, and you need to know when a run fails. On my board, finished work moves to done with proof attached. Every Friday I get a review of the week: what I got done, what I said I'd do and didn't, and anything that's been sitting for more than a week.
How long does it take to build an AI chief of staff?
Setting up a first version takes about twenty minutes: a Project with your context and your email and calendar connected. The 30-Day AI Starter Plan then adds one piece a week, a morning brief, one automated task, a daily close and written rules.
A team of agents takes longer. I've been adding mine one at a time since the spring, each one when a specific job kept falling through.
Do you need to code to build an AI chief of staff?
No. Steps one through eight run in a chat assistant in your browser with no code at all. Code only comes in if you want agents running on your own computer on a schedule, and even then you can describe what you want in plain words and have the AI write it.
I'm not an engineer. As I wrote in AI Fatigue Is Real, this past April the models crossed a threshold where a non-engineer like myself could suddenly begin coding. My system runs on Claude Code on my own computer now.
Should you build an AI chief of staff in Claude or ChatGPT?
Either works for the first version. Every prompt in the 30-Day Plan works word for word in both. Pick one and stick with it, because switching costs more than any difference between them.
I use Claude. My agents run in Claude Code on my Mac, with Notion as the record and my email, calendar and messages as things they can read.
What does it cost to run an AI chief of staff?
You can start for free. Claude's free plan includes Projects, connectors to your email and calendar, Skills and memory, which is enough for a first version. The paid Pro plan, about $20 a month, adds more usage and scheduled tasks.
My own system runs on paid tools, Claude Code and scheduled tasks, because it runs on its own every day. If you'd like to build yours with me, the live class is $99. If you want one built for your business, that starts with a free 30-minute call, and here's how my pricing works.
When should one assistant become a team of agents?
Split an AI chief of staff into several agents when one set of instructions covers jobs that conflict, when two jobs keep overwriting each other's work, or when some work needs to run on a schedule while you're away.
If none of that is happening yet, one good assistant with standing context, a morning brief and a daily close is plenty. Most people should stay there for months.
Want to talk it through first? Book a free 30-minute call. Or start on your own with the free 30-Day AI Starter Plan.