How to build an AI chief of staff with ChatGPT, Claude or Gemini
To build an AI chief of staff with a general AI tool, give it one narrow job: keep your list of open commitments, check it against the messages and email it can see, and draft the follow-ups for you to send. Put the instructions in a project in ChatGPT or Claude, or a skill in Gemini, connect read access only, and keep every send, accept and delete for yourself. Check its work every Friday. It works for one person.
Scope and date. Vendor features on this page are what OpenAI’s, Anthropic’s and Google’s own help pages said when we read them on 22 November 2026, as of November 2026; each is linked where it is used. They describe documented features, not a hands-on test of each tool. The starter prompt and routine are ours, written for this page, and are not a measured result.
Which page to read
For what the role is and a checklist for judging any product sold under the name, read What is an AI chief of staff?. For how it differs from asking an AI tool one question at a time, read AI chief of staff vs AI assistant. If you would rather buy than build, AI chief of staff tools lists the products sold under that name, in their own words.
The job to give it
A human chief of staff does the same few passes for one person, every day and every week. A general AI tool can do the reading and drafting part of those passes if you name them precisely and give it a list to start from.
Every working morning, starting from your list of open items, four lists:
- Unanswered requests. Someone asked you for something and has had no answer from you. Who, what, since when.
- Your commitments. What you said you would do, for whom, and the date, if one was said. Each item is marked as agreed, where you clearly said yes, or only mentioned, where it came up without a clear yes. That one distinction stops the list filling with things you never promised.
- Waiting on others. What other people said they would do for you, and how late each one is when a date was given.
- Drafts. A short follow-up in your words for each item that needs one, which you read, change and send yourself.
Once a week, one review: what you closed, what slipped, the commitments due next week, the items to renegotiate or drop with a draft message for each, and the decisions waiting on you with what is missing to make them.
Leave two things out of the job. Deciding which of two late items matters more is your judgement, not a reading task. Deciding who on the team owns a piece of work is the team’s decision. A setup that tries to do either produces confident lists that are wrong in ways that are hard to spot.
What it needs to see, and what it must never do on its own
| Source | Why it matters | How it gets in | Access to give |
|---|---|---|---|
| Your open list | The starting point of every pass, so nothing older than a few days drops off | Paste it, or keep it as a file in the project | None beyond the paste |
| Most external requests and promises | Paste threads, or connect your mailbox | Read and search. Drafting is useful; sending stays with you | |
| Team chat | Most internal requests and promises | Paste the threads, or connect the chat tool | Read the channels and direct messages you name |
| Calendar | Meetings where commitments are made, and dates you are already busy | Paste the week, or connect it where a read-only connection is available | Read only |
| Meeting notes | Promises made out loud | Paste them the same day | None beyond the paste |
Whatever the tool allows, these steps stay with you: sending, replying to or forwarding a message; accepting or declining an invitation; creating, moving or deleting a calendar event that other people see; deleting, archiving or sharing a file or message; and paying for anything. Write that list into the instructions, and set the tool to match where it lets you:
- ChatGPT. OpenAI’s app permissions page lists four levels. Always ask asks before reading or changing anything; Allow read actions reads without asking but still asks before making changes; Allow low-risk actions applies when no account, workspace or app policy overrides it; Allow all actions runs supported actions without prompts and may only be offered for an individual app or connected account. Which settings you see depends on your account, the app, the connected account and your workspace. Choose Always ask or Allow read actions for this job.
- Claude. The Gmail, Google Calendar and Google Drive connectors can send, reply to and forward email and create, update and delete calendar events; by default, outside Auto mode, Claude asks before each send, reply or forward. In Cowork, Anthropic describes three modes: Manual asks before actions, Auto keeps working with an automated check, and Skip neither asks nor checks. Use Manual. An Always allow choice you made earlier for a tool overrides the mode, so look in Customize, Connectors, at each connector’s Permissions tab for send, calendar and file tools, and set them to Always ask where that option appears. Anthropic says Always ask is rolling out gradually on paid plans, and on Team and Enterprise your organization’s settings come first.
- Gemini. Google’s Calendar help describes Gemini creating, editing and cancelling events and then showing a confirmation notification with Undo. It also says some things cannot be undone, such as an email sent to event guests when an event is changed. The page describes no approval step before the change and no read-only setting, so for this job paste your calendar instead of letting Gemini work in it.
Where to set it up in ChatGPT, Claude or Gemini
All three have a place for standing instructions, a way to connect email and calendar, and some form of scheduled run. The details differ.
| Feature | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Standing instructions | Custom instructions apply across your chats, up to 1,500 characters on Free and Go and 5,000 on Plus, Pro, Business, Enterprise and Education. Project instructions apply only inside that project and take precedence over custom instructions | Projects have their own instructions, chat history and knowledge files; available to all users, with up to five projects on a free account | Personal accounts, 18 and over: a skill, with a name, description, instructions and files. Work and school accounts: a Gem until Google moves them to skills (see below) |
| Reference files | Projects hold up to 5 files on Free, 25 on Go and Plus, 40 on Pro, Business, Enterprise and Edu | Project knowledge, with retrieval for larger collections on paid plans | Skills: files from your device or Google Drive, up to 100 MB in total |
| Email and calendar | Apps: Slack and Outlook email and calendar have their own setup pages, and scheduled tasks can use Gmail; availability depends on app, plan, region and workspace | Gmail, Google Calendar and Google Drive connectors for all users; on Team and Enterprise an owner must enable them first | Google Workspace, signed in with the same account; on work and school accounts an administrator must allow it |
| Scheduled runs | Scheduled tasks: up to 3 active on Free and Go, 5 on Plus, 10 on Business and Edu, 15 on Pro and Enterprise. Free runs at most once a day in a flexible window such as morning; exact times and hourly runs need a paid plan | Scheduled tasks in Cowork on paid plans (Pro, Max, Ultra, Team, Enterprise): hourly, daily, weekly or on weekdays. Cowork runs on desktop on all paid plans and on web and mobile on Pro, Max, Ultra and Team; Team and Enterprise admins control access | Scheduled actions, up to 10 active, rolling out gradually on personal accounts and available on qualifying Google Workspace editions |
Gemini: skills replace Gems on personal accounts
On personal Google accounts for people 18 and over, Google’s migration page says Gems became read-only on 17 November 2026 and are removed on 17 December 2026; Google moves existing Gems to skills automatically. To set this up as a skill, Google’s skills page gives these steps:
- On gemini.google.com, click Customize in the sidebar (under More if you do not see it), then Skills.
- Click Create, then Create manually.
- Enter a name in lowercase with hyphens, such as
daily-follow-up, a description, and the starter prompt below as the instructions. - Add files if you want, then click Create. Use the skill in a chat by typing
/and choosing it.
If you are moving an existing Gem, some files do not carry over: skills are capped at 100 MB, and audio, video, older Office formats such as .doc, and archives such as .zip are not supported. Download those from the Gem before 17 December and upload what you still need. On Google Workspace accounts, Gems stay until Google’s own timetable: March 2027 for business, enterprise and non-profit accounts and June 2027 for education accounts, so keep using a Gem there.
Three details that change the setup
- ChatGPT. A scheduled task created inside a project cannot open that project’s uploaded files, so put the full prompt in the task itself. Actions that send messages or change outside data may need your approval, which pauses the task until you look.
- Claude. A scheduled task that uses a tool set to Always ask may wait for your answer or finish without the tool. Tasks run remotely, even when your computer is asleep, unless you choose a local folder: then the task runs only while the computer is awake and Claude Desktop is open.
- Gemini. A scheduled action is prepared within the hour before its delivery time, so a message that arrives in that hour may not be in that morning’s lists.
To schedule the morning pass, create a task with an instruction such as: “Every weekday at 08:00, run the daily pass in my instructions and post the four lists here.” On ChatGPT Free, choose the morning window. Claude offers an on-weekdays frequency. Keep the Friday weekly pass manual, because it starts from the open list you paste.
The starter prompt
This is a starting point, written by us for this page and not taken from any vendor or course. Replace the parts in square brackets, cut what you do not need, and paste it into a project’s instructions, a skill or a Gem. It is about 2,300 characters, over ChatGPT’s 1,500-character custom-instructions limit on Free and Go, so use a project there. For a version built from your own answers, use the free AI chief of staff prompt builder, which also writes a daily check-in and a weekly review prompt.
Role
You act as chief of staff for [name], [role] at [company], a [number]-person
company. Your job is to make sure nothing [name] owes, or is owed, gets lost.
You prepare. [name] decides and sends.
Sources
Use as evidence only what I paste and these connected sources: [for example
Gmail, the #team channel]. Do not use earlier chats or saved personal
details as evidence. Start every pass by listing each source and whether
you could check it. If a source was unavailable, say so; do not treat
"nothing found" as "nothing happened". Do not guess names, dates or amounts.
Daily pass (when I write "daily")
Start from my open list: [paste or file name]. Reconcile it with messages
since the last pass that completed, then give me four lists:
1. Unanswered: requests to me with no reply from me. Who, what, since when,
and a short quote or link.
2. My commitments: what I said I would do, for whom, and the date if one
was stated. Mark each AGREED (I clearly said yes) or MENTIONED (no
clear yes).
3. Waiting on others: what others said they would do for me. Who, what,
and how late, only if a date was stated.
4. Drafts: for each item that needs a follow-up, a draft of two to four
sentences in my voice. Do not send anything.
Then: items to mark closed, each with the message that shows it is done,
and every item due in the next 3 working days.
Weekly pass (when I write "weekly")
Start from the open list I paste, then:
- What closed this week, with evidence, and what slipped.
- Open commitments for next week, by due date.
- Items to renegotiate or drop, with a draft message for each.
- Decisions waiting on me, and what is missing to make each one.
- Items with no update you could find, and the sources you checked.
Rules
- Never send, reply, forward, accept, decline, delete, archive, share, pay
or change a calendar event. Draft it and wait for me to do it.
- Quote or link the source for every item. If an owner or date is unclear,
write "unclear" and ask me.
- Never invent a due date. Calculate lateness only from a date someone
stated.
- Do not rank my priorities or assign work to other people.
- Leave out other people's private details that the follow-up does not
need, such as health or pay.
- One line per item, oldest first. If a list is empty, write "none".
After the first week, paste two or three of your own sent messages under a heading such as “How I write”, so the drafts sound like you. If it keeps marking things as agreed that you never agreed to, add one example of each kind from your own messages.
Your open list and the tool’s memory
The open list is the part that makes this work. Keep it yourself, as a short document or a file in the project, and update it after each pass: add the agreed items, remove the ones closed with evidence.
All three tools can also carry details between chats, and that is not the same as a list you can check. Look at the settings before you start:
- ChatGPT has saved memories and reference chat history, under Settings, Personalization, Memory, and a project can be set to project-only memory; see OpenAI’s memory help.
- Claude builds memory from chats by default on Free, Pro, Max and Ultra, and each project has its own; Settings, Your Claude shows what it holds, and you can turn off Use account memory on a project. See Anthropic’s memory help.
- Gemini can learn from past chats on personal accounts for people 18 and over, switched under Personal Intelligence in Settings; see Google’s help.
Whether you leave these on is your choice. The prompt above tells the tool to use only the named sources as evidence, so an item without a quote or link is a question, not a fact.
A routine, and a Friday check
Each working morning, about 10 minutes. Open the project, skill or Gem and type daily, or read the scheduled run. Read the line that lists sources checked: if one was unavailable, treat that morning’s lists as incomplete. Open the source of any item you are unsure about. Send the drafts you agree with yourself. Update your open list.
Each Friday, about 30 minutes, in this order:
- Sources. Read the sources lines from the week. Reconnect or paste anything that was unavailable.
- Recall. Pick one day. Open your sent email and chat for that day and find three things you promised. Check each one is on the open list.
- Accuracy. Pick three items and open their sources. Check the person, the date and the wording.
- False promises. Look for anything marked agreed that you never agreed to, and anything marked closed without evidence.
- Drafts. Count how many drafts you sent with only small edits. If you rewrote most of them, add your own messages as examples.
- Access. Open the tool’s connected apps. Remove any it did not need this week, and check that sending still needs your approval.
- Weekly pass. Paste the open list and type
weekly. Send the renegotiation drafts you agree with, and keep the updated list for Monday.
When to stop. If after two weeks the recall check still finds missing promises most days, look at where they were made. If they come from meetings, phone calls or an app the tool cannot read, more prompt work will not fix it: either paste those notes every day or use something that sits where those conversations happen.
Where the do-it-yourself route falls short
- It sees only the sources you give it. Promises made in a meeting, in a phone call, or in a messaging app you did not connect are not in its evidence.
- The open list is yours to keep. A chat is not a task list, and carried-over chat context is not a dated record you can check. If you stop updating the list, last week’s promise can drop out of this week’s pass.
- It acts for one person. Your mailbox does not contain what your colleagues promised each other. If ten people each run their own prompt, the company has ten private lists that nobody else can see.
- Connected access is broad. A mailbox connection can usually read the whole mailbox, not only the threads this job needs. Decide what you are comfortable with before you connect it, and check your company’s rules; the AI acceptable use policy template is a place to write them down.
- Scheduled runs depend on your plan. On the pages we read, Claude’s scheduled tasks in Cowork need a paid plan, ChatGPT Free runs a task at most once a day in a flexible window, and Gemini’s scheduled actions are still rolling out on personal accounts.
When a product makes more sense
Building it yourself is a good first step for one person, and a fair test of whether the job is worth doing at all. A product built for the role makes more sense when:
- The commitments are between people on the team, not only to you, and the person who owns a task needs to see it where they work.
- The daily paste is the step you skip. If the sources have to be copied in by hand, the list is only as current as the last time you did it.
- Access should follow the team’s rules, such as who is in which conversation, rather than one person’s connected account.
- You want the list kept as tasks with dates, in a place the team already looks, rather than in a document you update by hand.
Whatever you choose, ask the questions in the evaluation checklist: what it reads, what it keeps, and which actions wait for you.
Where Opitor fits
This section describes our own product. Opitor offers direct and group messages, channels, and a Todo List that tracks tasks and deadlines. Each person has a personal AI chief of staff in Messages to help with messages, todos, deadlines and drafts. It looks up only what you have access to, proposes todos for you to confirm, and AI replies are drafts you choose to send, signed in purple. Opitor tracks tasks and deadlines in Todo List, beside the conversations they come from, which fits a team that wants shared follow-up next to its messages rather than one person’s private list.
Gmail is separate. In Settings, under Personal connections, you can connect your personal Gmail to your personal space, separately from Google sign-in, and read your recent Gmail messages in Opitor. The connection is read-only: Opitor cannot send, change or delete anything in your Gmail.
What it does not do: Calendar shows accessible todo and subtask deadlines; it is not a meeting scheduler. Administrators cannot read other people’s private messages or groups they have not joined. Opitor has no paid plans currently. The FAQ lists what is included, and AI chief of staff tools sets Opitor beside other products under the same name.
Sources
- OpenAI: custom instructions, projects, apps, app permissions, scheduled tasks and memory.
- Anthropic: projects, Google Workspace connectors, Cowork modes and permissions, scheduled tasks in Cowork and chat search and memory.
- Google: skills, Gems to skills migration, Gems, Google Workspace in Gemini, Google Calendar in Gemini, scheduled actions and memory.