Build your own AI Chief of Staff
Michelle Hamilton's grand return. Stop re-explaining yourself to your AI every morning, give it two files it can carry anywhere, and hire it the way you would hire a person: one task at a time.
Everybody in this group has the same wish and the same objection. The wish is a chief of staff: somebody who already knows what today looks like, what is about to slip, and who has been waiting on you longest. The objection is that you do not have one, you are not going to be given one, and you are not sure you want an algorithm anywhere near your calendar. Michelle's answer was to introduce hers.
She presented from Naples, Florida, mid-vacation, driving to Miami for CREW Convention on Tuesday. So the demo she ran was the one where her assistant tells her she is on PTO. That turned out to be the better demo.
Meet George
Michelle's chief of staff is built on Claude Cowork and is called George. She asked him live, on the call, to explain himself. Lightly condensed, he said this:
"My job is to reduce the number of decisions Michelle makes in a day." Three priorities, ranked. First, prepare her for her meetings. Second, make sure her email is actually handled, not just sitting there. Third, keep a task list she trusts, where the repeatable work is already mostly done before she looks at it. Connected to her calendar, email, Google Drive, Slack and meeting transcripts. Every morning he pulls all of it into a six line brief. Nothing more, because the point is cutting down what she has to process, not adding to it.
Those six lines are worth writing down, because they are the actual specification. What is on the calendar. The one thing that absolutely cannot slip today. The one thing that will slip if she does not touch it. Who has been waiting longest on her. The one decision only she can make. What got taken care of overnight.
Notice what is not in there: a summary of everything. A generic assistant tells you that you have four meetings tomorrow. You had a calendar for that.
The part people expect to hear and did not
Michelle does not let her AI send email, put things on her calendar, or reply in Slack on her behalf. It drafts, into her drafts folder, and she signs off. Her reasoning was not fear. It was that an email from her should sound like it came from her, and that the volume is the problem worth solving, not the sending.
So the ceiling on this whole build is lower and more useful than the pitch usually implies. You are not automating yourself away. You are arriving at things already prepared.
What you should be able to do by Monday
Five capabilities, in order. Each one is load bearing for the next, and the last one is the only one most people skip.
Know what your AI can already reach
Open a new conversation on a light, fast model and ask it to list every external source it currently has access to. Not what you think it has. What it says it has.
You've got it when: you can name the connectors that are live, the ones that are missing, and which of the missing ones your employer has actually blocked rather than simply not turned on.
Let it interview you, and answer honestly
The model does not know your role, your industry, your reporting lines, your travel, or who in your house has practice on Thursday. It asks; you answer; where something is unclear it says so instead of inventing a plausible version of your life.
You've got it when: the interview has asked you at least one question you had to stop and think about.
Walk away with two files that belong to you
A profile file and a working rules file, in Markdown, which is just a plain text document with light formatting. They are yours, they are portable, and they are the end of explaining yourself to a blank chat window.
You've got it when: you can open both files, read them, and correct a line yourself without asking the AI to do it for you.
Set the boundary before you need it
Work the four Ds: what you delegate, how you describe it, what it may decide on its own, and where it must stop and ask. Draft first. Act later.
You've got it when: you can name, out loud, the one action it is never allowed to take without you. Most people's is "send."
Correct the first answer instead of admiring it
Run "prepare me for tomorrow," then tell it what was useful, what was noise, what it assumed and got wrong, and what it missed entirely. That correction is the training. There is no other training.
You've got it when: a week in, the brief tells you something you did not already know, and you have stopped having to ask for the same thing twice.
Chat or Cowork, and why it matters
Michelle's rule of thumb: you talk and ideate in chat, one task at a time. You move to Cowork when the job needs several places at once. Cowork is the room full of coworkers. Building the profile through an interview is a conversation, so it stays in chat. "Prepare me for tomorrow," pulling calendar and email and drive together, is a multi-source job, so it goes to Cowork on a mid-weight model.
The same split exists in the other platforms under different names, which is genuinely confusing and not your fault. ChatGPT calls its version chat and work. Copilot is its own animal with its own instruction set. The naming is a mess across all three; the distinction underneath it is real.
Model choice is a cost and footprint decision
For quick, simple questions, use the small fast model. It is cheaper in tokens for you or your company, it answers faster, and it draws less from the data center. Michelle also runs her tools on load-when-needed rather than always-on, for the same reason: a request for a spaghetti recipe should not be loading your entire board history.
Projects, or the intelligent filing cabinet
Build the whole thing inside a project rather than a loose chat. A project is a drawer that remembers everything you put in it, does not wander outside itself, and can be updated by dropping in a new file instead of hunting through six weeks of old conversations for the place you explained something. Your two Markdown files live there permanently, which is what makes every future conversation start from context instead of from zero.
That line was Michelle's, and it arrived in the middle of the most practical stretch of the hour. Jodi Tagessen, on the CREW board, took it a step further: if your company is writing an AI policy, the first question is not Claude or Copilot. The first question is which points you actually care about, human in the loop being the obvious one, and only then which tool you allow and what it may connect to.
And then the part that is easy to skip past: your employer's permission and your own comfort are two different settings. They may let it send email on your behalf. You are still allowed to decide you would rather it did not.
Watch it back.
The full hour, including the fifteen minutes of questions that broke out halfway through and were better than the planned material.
The full walkthrough, the prompts and the FAQ are already below and on Michelle's guide. Trouble opening the video? Email Ivy and she will sort your access.
Seven moves, in this order.
Michelle built hers over months. This is the version you can start on a Saturday morning. Do not skip step one and do not do step seven only once.
Pick a lane: work or personal
Both are legitimate and there is no right answer. Work means meetings, follow-ups, what happened while you were out, and it means living inside whatever your employer has approved. Personal means who has practice tonight, did that hotel ever get booked, and what should I read at the airport. If your company is restrictive, build the personal one first. You will learn exactly the same skills on data nobody has a policy about.
Test what it can already reach
New conversation, small fast model, one question: what external sources can you access right now? Michelle ran this live and her own Claude reported back that in that conversation it did not have Google Calendar loaded, which was the point. If something is missing, the plus button and then Connectors shows you the list. If a connector you want is not there, that is where step three begins.
Find out what is blocked, and accept the answer
A connector that will not connect is usually not broken. It is usually your IT department, and they had reasons. Ask them plainly which tools, connectors and read or write actions are approved for your account. Michelle's suggestion, which is a good one: send them the prompts you intend to use, so the conversation is about something specific rather than about "AI." Ivy's own situation is the illustration. EyeQ runs Claude Enterprise, and will still not let her connect the Microsoft ecosystem to it. That is the rule, and she follows it.
Create the project before you create anything else
Projects live in the left-hand rail in both Claude and ChatGPT. Name it, open it, and do the entire build inside it. Everything you are about to produce becomes permanent context for every future conversation in that drawer, which is the whole reason this works. Start the interview in chat mode inside the project.
Let it interview you
This is the step that feels slow and is not. It asks about your role, your industry, your responsibilities, who you are accountable to, your hours, your travel, how you like to be communicated with. Edit the prompt before you send it: strip "industry" if this is for your family, add questions about a multi-generational household or a three-child logistics problem if that is your actual life. And instruct it plainly not to invent answers, to ask when something is unclear, and to draft a concise working profile for your review at the end. You are the expert here. It is taking dictation.
Produce the two Markdown files
Staying in the same chat, ask for two downloadable files: Your_Name_Chief_of_Staff_Profile.md and Chief_of_Staff_Working_Rules.md. Markdown is a plain text document with light formatting. That is the entire mystery. It matters because models read it fast and cheaply, and because a file you can open in any text editor is a file you own. Update it when life changes, a promotion, a graduation, a new territory, by uploading it and saying what moved. Upload both into the project and they become permanent.
Run it, then correct it, then correct it again
Ask for tomorrow. Read what comes back. Then tell it, specifically, what was useful, what was noise, what it assumed and got wrong or could not support, and what it missed. That is the job. Michelle's framing is the one to keep: you would not hand a brand new employee twelve tasks on day one and expect all twelve done perfectly in five minutes, and then be surprised when they guessed. Guessing is what a hallucination is. Most of them trace back to a question asked badly and a model that was never made to say where it got something.
And then schedule it
Once the brief is good, it can run on its own. Michelle's lands in Slack at 7:00 AM every morning. Yours could be a nudge every three hours that says which priority clients have emailed you. Schedule what earns its place and nothing else: a job that runs because it can rather than because you need it is just cost with a calendar entry.
Her longer arc, one to steal: give it a week of daily briefs, then on Friday ask it to review itself. How do you think that week went. What did you not understand. Where did I correct you. What should I have told you at the start that would have made this easier. It is a performance review for an employee who cannot be offended by one.
Connected does not mean approved.
Five rules Michelle applies before any assistant of hers touches anything. None of them are technical. All five are the ones that get skipped.
Connected is not approved
That a tool can reach something does not mean your organization has agreed that it should. Those are two separate questions and only one of them is answered by a green check mark.
Personal accounts stay personal
Do not move company information into a personal account because the interface is nicer. Use the environment your employer approved for that information.
Training and retention differ
Whether a provider trains on your data and how long it keeps that data are different policies with different answers. Know both for the account you are actually using.
Grant the minimum
One connector that solves a real recurring problem beats six that look impressive. Earn the next one.
Draft first, act later
Let it find, summarize, organize, prepare, recommend and draft. Be deliberate before you let it send, delete, publish, schedule, purchase or commit you to anything.
Delegation, description, discernment, diligence.
Michelle frames the governance conversation with your assistant as four questions, and has it interview you through them the same way it interviewed you about your job.
Delegation: what goes to the model and what stays with you. Description: the who, what, where, when, why and how of the task, because a task described badly is a task done badly. Discernment: what it may decide on its own versus what it must ask permission for. Diligence: where it must stop, full stop.
The output of that conversation is the working rules file. Which means your boundaries are written down, portable, and reviewable, instead of living in your head and getting relitigated every time you are in a hurry.
"If you wouldn't put it on LinkedIn, it doesn't belong in your AI."
Michelle's blunt version, with the caveat she added herself: the platforms already know an enormous amount about you. Facebook and Instagram know her children's names and her grandson's name and who she works for, and she put that there. What she would not hand them is her bank details or a card number, and the same line applies here.
Her Markdown prompts were deliberately written not to request private company data, HIPAA-protected information or personally identifiable information, so that you can show them to IT and have a short conversation rather than a long one. That does not make it their approval. It makes it an easier ask.
Nothing on this page is legal advice, and platform policies change. Verify against your own organization's governance before you connect anything to anything.
Steal these.
Michelle's own copy-ready prompts live on her guide, linked below, with per-platform setup for Claude, ChatGPT and Copilot. These are the AI × CRE working versions of the same sequence, written so you can run the whole build from this page. Edit them before you send them. They are a starting point, not scripture.
What the room asked.
The chat ran hot for most of the hour and Michelle offered to answer the overflow in writing. She did. Product details were checked September 22, 2026, and these things change weekly, so treat every answer as true on that date and verify before you rely on it.
You do not start from zero. Ask your current AI to write your Chief of Staff Profile and Working Rules from what it already knows about you, then keep those files as your portable source of truth. You can read them, edit them, and use them anywhere.
Claude also now supports importing memory from other AI providers on supported plans. Think portable context, not a perfect copy of one account into another.
Not perfectly. Chats, memory, projects, files, permissions and settings are stored differently everywhere, and an export is not a recreation of the original account. ChatGPT lets eligible users export conversation data; Claude has a built-in memory import flow.
Your two files travel easily, which is the point. The goal was never to move every conversation. It is to make the context that matters portable.
Possibly, yes. If Outlook, Teams, OneDrive and SharePoint are where your work already is, your organization has probably already settled identity, security, permissions and data access there, and Copilot Notebooks can hold files, references, conversations and instructions in one persistent workspace.
Start where your work already lives and where your company allows you to work. Do not move company information somewhere else because you prefer the interface.
One question: where does the information this task needs already live?
- Outlook, Teams, OneDrive or SharePoint, start with Copilot.
- A Claude project, Claude connectors, or work already organized in Claude, start with Claude.
- A ChatGPT project or a workflow already running there, start with ChatGPT.
Your working rules stay the same even when the platform does not.
Yes, and it is not bypassing IT. Use the AI environment your organization approved for that information, keep company data inside approved systems, and use personal accounts only for what your company permits there.
Ask IT exactly which tools, connectors and read or write actions are approved for your account. A blocked connector is a governance boundary, not a technology failure.
They are much less technical than they sound. A Markdown file is a lightweight text document. That is all.
- Your_Name_Chief_of_Staff_Profile.md: who you are and how you work.
- Chief_of_Staff_Working_Rules.md: how you want your chief of staff to behave.
Simple, readable, fast for AI to use, easy to move. Files that belong to you rather than to one platform.
Maybe not. Anthropic is combining the two experiences for some plans. If you still see both, use chat for conversation and Cowork for larger multi-step work. If the selector has vanished, nothing is broken; ask for what you need and Claude decides whether it is a quick answer or a bigger job.
Team and Enterprise environments may still show them separately depending on rollout and admin settings.
No. The web version is enough for a cloud-based chief of staff: projects, connected cloud services, research, cloud files.
The desktop app earns its place when you want local files, local connectors, browser use, or computer-level workflows. Michelle uses hers with access limited to specific folders, which is called sandboxing.
No. Please do not. Start with one useful connection. Calendar, and "prepare me for tomorrow." Then perhaps email, and "what am I waiting on, and who is waiting on me?"
Add the next connection only when it solves a real recurring problem. Earn the next connection.
Yes, and several people in the room already do. ChatGPT Shared Projects are the practical option: common files, instructions and conversations, with project-only memory, so they do not pull in a member's private context from outside the project. Travel, caregiving, home projects, college planning, schedules.
Shared means shared. Be thoughtful about what goes in.
Copilot Notebooks can be shared in supported organizational environments. Microsoft currently says Notebook sharing is not yet available for Microsoft 365 Personal, Family and Premium subscribers.
Start with less than you think. Let it find, summarize, organize, prepare, recommend and draft. Be far more deliberate before you let it send, delete, move, publish, schedule, purchase or commit you to something.
Draft first, act later. The human in the loop is not a limitation, it is the design.
It happens. A model can produce sophisticated work and then confidently get a basic date wrong. Ask it to verify anything date-sensitive, give it access to a current source, ask it to show where the answer came from, and check the original yourself when the consequence matters.
The polish of an answer is not proof that it is right.
Anthropic: why Claude gives incorrect or misleading answers ↗
That one was just for fun, but yes, Olive is real, and she is modeled on Michelle's actual dog. On supported ChatGPT desktop experiences: Settings, then Pets. Pick a built-in or custom pet. The floating controls also give you quick ChatGPT access without opening the full window.
Availability varies by plan, app version and workspace settings.
Chatting is ask a question, get an answer. Agentic work is giving the AI a goal, context, tools, boundaries and a multi-step job to complete, across files and systems, on a schedule.
You do not need to race there. But learning the difference now puts you well past getting better at prompts.
One attendee named it exactly: constructive struggle. The goalposts do keep moving. The first answer will not be right, you will correct it, you will teach it what matters, you will strip out what does not, and the thing gets better over months.
You do not need to master AI. You need to get comfortable learning it.
Three things the group needs from you.
The postmortem
Sarah's idea, and the room liked it: everyone builds a chief of staff from these materials, then we come back and run a diagnostic. What worked, what broke, where you got stuck, what you would tell somebody starting today. Tell us if you want it and it gets scheduled.
Vote for it →Member spotlight
Five minutes at the top of a session for one member to show a win or an aha moment from the past month. A tool, a prompt, a workflow that finally clicked. It has been on the agenda for months and still nobody has claimed it. Be the person.
Claim a slot →Tell us what to cover
Every text box on this site feeds a form the working group and board actually read. Speakers you want, topics you want covered, prompts you are using, a question you would rather ask in writing. Drop it in and it gets routed.
Open the survey →Session 08: AI Without the Anxiety. Anthony Peters, IT Systems Manager at BranchPattern, on security, data boundaries, permissions, grounding, hallucinations and who is accountable for what, in plain language. Friday October 16, 10:00 AM ET. Register on Teams.
Michelle flagged two future sessions of her own: turning exported chat history into reusable skills, and attunement, which is teaching a model how you prefer to think and communicate. She is also now certified to run Anthropic small business training and plans to bring one to this group. Dates to follow.
The leave-behind.
Michelle's full interactive guide with per-platform setup, plus the starter files and the prompt pack. Free, no gate, no form in front of it.
Build Your Own Chief of Staff
Michelle's own interactive guide, hosted by AnswerRocket. Three lanes for Claude, ChatGPT and Microsoft 365 Copilot, with copy-ready prompts and post-prompt checks at every step. This is the primary resource from the session.
Open the guide FAQ · On this pageThe sixteen questions
Everything the chat asked and Michelle answered in writing afterwards, with links to the current product documentation for each. Checked September 22, 2026.
Jump to the FAQ Starter file · MarkdownProfile template
A blank version of the first file, with the sections the interview is going to fill in. Rename it with your own name, open it in any text editor, and edit it by hand whenever something changes.
Download .md Starter file · MarkdownWorking rules template
The second file, structured around the four Ds. What it may draft, what it may decide, what it must ask about, and the one thing it may never do without you.
Download .md Prompt Pack · LibraryEvery session's prompts
This session's seven are on this page above. The library is where all of them live together, from Session 3 onward.
Open library Recording · VideoWatch the session
The full hour, including George introducing himself live and the fifteen minutes of questions that broke out halfway through. Hosted on SharePoint.
Watch recordingMichelle leads AnswerRocket's AI adoption and change management practice, training teams on Copilot and Claude around the world, and recently certified with Anthropic as a small business trainer. She is a returning presenter here; her AI 101 sessions are where a lot of this group started. Her position, repeated often enough that it is clearly not a disclaimer: these are portable skills you are building for yourself, not for an employer, and you take them with you.