Your Business OS kit

• An AI you pay for. Claude or ChatGPT on a paid plan. Both can read a folder and both can run scheduled tasks. Use the one your company already approved.
• Obsidian, free, from obsidian.md. It is a window onto a folder. It is optional today and you will want it by next week.
• An empty folder on your laptop, named after your business. Keep it out of a synced drive for now; we will talk about that in Stage 6.
• This page in a second tab.

Memento Man, from the talk: the model is sharp, fast and forgets you every morning. Leonard's fix was notes outside his head. Yours is a folder.

Plain text files, because:

• Any model can read them. Claude today, something else next quarter, a local model for the data you cannot send out.
• You can read them. No vendor's memory feature shows you what it decided to remember.
• They have history. Obsidian keeps file history; git keeps all of it. You can ask "what changed this week?"

Mine, top level: Identity, _Co-CEO (the rulebook and open threads), Clients, Prospects, People, Meetings, Reference, Daily. Yours will look different. Start with Identity and one folder for the accounts or projects you care about.

The model writes your first five files by interviewing you. This is the one prompt worth pasting, because the answers are yours.

Open a new chat (a Claude Project or a ChatGPT Project, with your folder attached if your plan allows). Paste:

You are setting up my Personal OS: a folder of plain text files that any AI model will read before helping me. Interview me, one question at a time. Push on vague answers. Ask about:
1. Who I am: name, role, company stage, who I sell to and who I do not.
2. Where I am headed: the one goal that decides my choices right now, then this quarter and this year, with numbers where I have them.
3. What drives me: the values I will not trade, and who I want to help most.
4. What I protect: the days, hours and people that are off limits, and what needs my approval before anything is sent or changed.
5. What I am bad at and want help with. Where I say yes when I should say no. What I undervalue about my own work.
6. More of, less of: the work that gives me energy, and the work that drains me even when it pays.
7. The tensions I am holding: the two things I want that pull against each other.
8. What I want you to help with: your job description, in my words.
9. The three accounts or projects that matter most right now.
Then produce five Markdown files: identity.md, goals.md, voice.md, how-i-work.md, operating-agreement.md. Put questions 5 through 8 in identity.md under their own headings; that is the part that lets you push back on me later.
Mark anything you inferred rather than heard as [inferred]. Keep each file under 300 words.

Answer in three sentences per question, and answer 5 through 8 like you would tell a friend, because those are the ones that make the brain useful. A brain that knows your goals writes a decent brief. A brain that knows what you are bad at tells you when you are doing it again. Save the five files into your folder (Obsidian: new note, paste, name it). Then read them and find every [inferred]. Fix those lines in the file. That is your first corrections pass, and it is how the system starts to sound like you.

A brain with only an identity file is a profile. The value shows up when the pages about your work are in it. This stage is judgment, so there is no prompt. Three questions place anything:

Conventions that keep pages trustworthy (mine; use them or write your own)

• Every account or project page starts with the same four fields: status, next step, last touch, updated. Your scheduled tasks read those fields, so they have to exist.
• Below that, Recent activity: one dated bullet per development, newest first, capped around ten.
• One file, one home. Decks and contracts live where they live; the page points to them.
• Credentials never go in. Name the folder they live in and say it is off limits.

WORKED EXAMPLE

operating-agreement.md came out of the interview as a draft. Make it real. Mine has five sections and fits on one page:

1. Order of authority. When instructions conflict: my current request, this agreement, the how-it-works notes, area instructions, the model's own habits last.
2. What needs my approval. Sending anything. Changing the CRM. Publishing, purchasing, deleting. Reading and analyzing never needs approval.
3. Where things live. Thinking layer (this folder), deliverables layer (the other folder), off limits (credentials).
4. How to write to the brain. Four fields on every page, Recent activity newest first, link people and accounts.
5. What never happens. Nothing auto-deleted, nothing moved without a yes, no bulk push to the CRM.

Then point every tool at it. A Claude Project's instructions, a ChatGPT Project's instructions, a CLAUDE.md in the folder: each is three lines that say "read operating-agreement.md first." Change a rule once and every tool follows it.

A brain you build once is an intranet nobody trusts. What keeps it alive is a few scheduled tasks that read it and update it on their own. The easiest first one is the morning brief, and you are going to design it, because the design is the lesson.

Every scheduled task has four parts. Fill these in before you set one up:

Then write the instructions from your table. Starter, copy and fill in the blanks:

Every weekday at [time], read [files]. Give me [what it hands me], in under [length]. Then update my files: [what it changes]. Mark any guess as [inferred]. Never delete anything. If a file you need is missing, say so instead of guessing.

Set it up: in Claude it is called Scheduled tasks; in ChatGPT it is Tasks. Point it at your project. Run it once by hand now.

Read the output. Fix one wrong thing in the folder, never in the chat. A correction in the chat is gone tomorrow. A correction in the file is permanent. Run it again.

Mine found about 100 contradictions in six weeks while every scheduled task ran on time. The rules that came out of that, so you can skip the six weeks:

1. Every update also fixes what it just made out of date. When the task adds "proposal sent," it also changes the next step that said "send proposal."
2. If an update waits for you to approve it, it never happens. Let the task write the facts now, mark its guesses, and you correct them.
3. One file, one home. Every document lives in one place, and the brain points to it.
4. A missing field cannot go stale. If a page has no "next step" line, nothing will ever flag it. Make sure the line exists before you worry about whether it is current.
5. Instructions to future sessions are the most dangerous content. "Do not flag this again" keeps a stale fact alive forever.

Two data questions to answer before the first transcript lands: where does customer personal data go (summary only, scrubbed locally, or a private instance), and is your folder in a synced drive that other tools can see. Decide on purpose.

Personal is rung one. Departmental moves context, memory and skills into a shared folder: one voice page, one ICP page, shared procedures, and the contractor who starts Tuesday inherits them. Business OS is every department reading the same customer record from signed SOW on. Three stacking rules: each fact lives once in the layer that owns it; agents read bottom-up, company then team then you; edit rights tighten as you go down.

Write your 30-day plan on one line:

By ____ my folder has ____ pages, a scheduled task runs every ____, and ____ moves to the team folder.

Send me the first brief your scheduled task produces. I read every one, and I will tell you what it guessed.