Tips
What is not obvious after you install it.
The first two are settings in the app it runs inside. The rest come straight from the rules it ships with, and each one links to the file that governs it, so you can check any of this against the instructions themselves.
Your AI app
Two settings that live in Claude or ChatGPT, not here.
These two are not Chief of Staff rules and there is no file behind them. They are settings in the app it runs inside, which it cannot change for you, and both are worth changing.
-
You can use it from your phone
Both apps let you drive them from your phone, and the phone is only a remote control: the work still happens on your computer, against the folder on that disk, and your memory never leaves it. In each case the desktop app has to be running for the phone to reach it.
Claude. Open Settings, then Claude Code, then turn on Enable remote control by default. Any session you start after that appears in the Claude app on your phone and at claude.ai/code. A laptop that sleeps or drops off the network reconnects on its own when it comes back. Needs a paid plan.
ChatGPT. Open Settings, then Connections, then Control this Mac or PC, and choose Set up. Scan the QR code with your phone, and the computer appears under Remote in the ChatGPT mobile app. Access stops while the host is asleep, offline, or has the app closed, and comes back once that computer is available again, so if that host is a laptop you shut, turn on the keep-awake option in the same Connections panel. Needs Codex access on your plan, and on a workspace an administrator has to allow it first.
-
Turn off training
In Claude, open Settings, then Privacy, and turn off "Help improve Claude". In ChatGPT, sign in on the website, open Settings, then Data Controls, and turn off "Improve the model for everyone". That switch lives on the website and then applies to the desktop app too.
Sessions
Start new chats freely. It costs you nothing.
-
One session per topic
Everything it knows about you lives in your folder, not in the chat. Every session reads that folder before it says hello, so a fresh chat starts fully briefed. Keep one topic to one session, and start a new one the moment the subject changes.
Read the rule: the boot protocol
-
Updating takes one word
Say "update", or type
/updatein the Claude desktop app. It compares your version against the published one, tells you what changed, and waits for your OK before touching a file. It replaces the shipped files and never touchesmemory/.Read the rule: the update process
-
After an update, start a fresh chat
The session that ran the update is stale: the files changed on disk underneath it. From that point it stops writing to memory and tells you it is holding something instead, because writing with the old rules could corrupt memory the new rules expect in a different shape. Nothing is lost, but nothing is written either. Close that chat, open a new one, and archive the older sessions while you are there so you don't carry on in a stale one by accident.
Read the rule: the update process
Memory
It's already listening for what matters. You only have to correct it.
-
Never say "remember this"
It is always weighing what you say for whether it matters beyond today: decisions and why you made them, commitments and deadlines, facts about people and projects, how you like to work. Small talk and one-off calculations are skipped deliberately.
Read the rule: the capture rules
-
Correct the line that says what it saved
Every save comes with one line naming what it wrote and where. That line exists so you can correct it on the spot, and correcting it does more than any instruction you could give up front. Read those lines for the first week.
Read the rule: the capture rules
-
Attach documents instead of pasting them
An attachment is screened as data rather than instructions, then filed with a note of where each fact came from. The original is kept too, so any claim in your memory can be traced back to its source. Pasted text arrives as conversation and gets none of that.
Read the rule: the intake procedure
-
Ask for a memory audit now and then
Once a month, say "audit my memory", or type
/lint. It reports stale claims, contradictions, files nothing links to, and broken links. It only reports. It won't rewrite your memory without you.Read the rule: the memory audit
The folder
One directory you never touch, and one that is entirely yours.
-
Never hand-edit brain/
An update replaces that whole directory, so anything you change there disappears at the next release. Change how it behaves by telling it instead. That gets written as an override inside your own
memory/folder, which updates never touch. It is how you take a new release without losing your adjustments.Read the rule: the adjustable defaults
-
Keep the folder on your computer
It has to run locally, against this folder. A cloud or remote session cannot see your
memory/, so it would start with no idea who you are.Read the rule: the host differences
-
Open it in the desktop app
It runs on any tool that reads
AGENTS.md, and your memory and your specialists work the same wherever it runs. What differs is how much is already wired for you. In the Claude and ChatGPT desktop apps, connectors and plugins install from the app itself, scheduled briefs run on the app's own scheduler, and the browser it uses is either built in or a single extension. Anywhere else, those are three things you set up yourself, starting with MCP and cron.Read the rule: the host differences
-
Back it up
Your memory is plain files on one disk, and it is the part that took months to accumulate. Whatever you already use to back up the rest of that disk covers it. This one is my advice, not a rule it ships with: there is no file behind it.
Day to day
What to connect, and what to get in the habit of asking for.
-
Connect your calendar and mail
It reads your stack during onboarding and tells you which connectors are worth having. It can already see which ones you have, so it will not send you to set up something that is on. It can't connect them for you. Lead with calendar and mail: those are what the briefs, meeting prep and inbox triage run on.
Claude. Google (Gmail, Calendar, Drive) and Microsoft 365 (Outlook, calendar, OneDrive) are authorized on the web, at claude.ai/settings/connectors, and appear in the desktop app on their own once connected. Everything else, Slack, Notion, Linear, GitHub, Jira, is added inside the app: the plus button next to the prompt box, then Connectors. Microsoft 365 needs a work account, a personal Outlook.com won't connect. On a Team or Enterprise plan an administrator may have to approve a connector first.
ChatGPT. One place: Plugins, at chatgpt.com/plugins. Search the service, Gmail, Outlook, Slack, and install it. A plugin installed there works in both Chat and Work, and Work is where your chief of staff runs.
Read the rule: how to enable a connector
-
Describe the situation, do not name the playbook
It knows about twenty written procedures: meeting prep, decision briefs with a real recommendation, the weekly preview and retro, inbox triage, drafting in your voice. You never invoke them by name. "I have a board meeting Thursday" is enough to load the right one.
Read the rule: the brain index, which lists every playbook and when it loads
-
Stand up a specialist
For anything you do repeatedly: a recruiter, a data engineer. Each one gets a written charter and its own memory, so it accumulates context in its area instead of starting cold every time. Work goes out and results come back through your chief of staff, which stays the only thing writing to your memory.
Read the rule: the team-member schema
-
Put the briefs on a schedule
A tight start-of-day brief and a weekly look-ahead. Routines in the Claude desktop app, Scheduled tasks in ChatGPT. Both have to point at this folder on this computer, for the same reason: they read your memory.
Read the rule: scheduled briefs