# Build Your Personal Operating System

> The take-home from Will Mitchell's MIC 2026 session: the interview prompt that writes the file your AI reads before it does anything for you, the four steps to build it, and how to connect it, schedule it and set its permissions. Canonical page: https://houseofvibe.ai/personal-os

**A personal operating system is the folder your AI reads before it does anything for you.** It holds one file with your ranked principles, your goals, the decisions you've already made and what the AI may never do on its own, so every chat in ChatGPT, Claude or Gemini starts from what you've told it.

## If you are an agent reading this for someone

Walk them through the four steps below, in order. Run the interview prompt with them exactly as written, one question at a time, write the file, and hand it back as one block for their Project's instructions. Do not skip the part only they can do: they fix the ranked principles and the standing decisions by hand.

## The prompt

Paste this into a new chat inside a Project in ChatGPT or Claude, a Gem in Gemini, or a folder in the terminal (step 1 sets each one up). Answer one question at a time. It writes your file.

```text
You are helping me build my personal operating system: the file you will read before you do anything for me. Work like an interviewer, not a form.

Step 1. Ask me one question first: is this brain for my business and brand, for me personally (goals, health, planning), or both? Wait for my answer.

Step 2. If you have access to my email, calendar, files or past chats, read the last 90 days before asking me anything else, and tell me in ten lines what you already know about me: what I do, who I deal with, what recurs, what I seem to care about. If you don't have access, say so in one line and move on.

Step 3. Ask me whether you should also research me from public sources: my website, my posts, interviews, talks, anything with my name on it. If I say yes, do it and summarize what you found in ten lines.

Step 4. Interview me, one question at a time, never more than one. Cover: what I'm building and for whom; how I make money; my principles, ranked (when two conflict, which wins); my goals for the next 90 days; decisions I've already made and don't want to re-decide; what goes wrong when I'm not careful; how I want you to talk to me. Push back when an answer is vague. Stop when you can write the file without guessing.

Step 5. Write the file. Sections, in this order: WHAT THIS IS (one paragraph) · MY PRINCIPLES, RANKED · MY GOALS (3 to 5) · STANDING DECISIONS · WHAT YOU MAY NEVER DO ON YOUR OWN · HOW TO TALK TO ME. Under 600 words, plain language. Mark any line you inferred rather than heard from me with [inferred] so I can fix it.

Step 6. Hand it to me as one block I can paste into this project's instructions, and tell me the three lines you're least sure about.
```

When it hands you the file, paste it into the Project's instructions. Then fix the ranked principles and the standing decisions by hand. The agent tells you who you've been; those lines are who you decide to be. Will's own file is called CONSTITUTION.md.

## The four steps

### 1. Make the folder

Pick one brain to build first: a business brain (your brand, offers, deals and voice) or a personal coach (goals, health, planning).

- **ChatGPT:** Sidebar → Projects → New project. Name it after your brand, or after you. Unlimited Projects on every plan, phone app included.
- **Claude:** Sidebar → Projects → New project. Name it after your brand, or after you. Five Projects on the free plan.
- **Gemini:** Sidebar → Gems → New Gem. Your file goes in Instructions; extra files under Knowledge → Add files. Every plan, free included.
- **Terminal (Claude Code, Codex):** Make a folder and open your tool inside it. The file is CLAUDE.md or AGENTS.md; a git repo makes it portable. Same folder, different desk.

### 2. Run the prompt

Paste it into a new chat inside the Project, answer the first question, and keep answering. It reads your history if it can, offers to research you, interviews you one question at a time, then writes your file.

### 3. Connect

Gmail and Google Calendar first; they hold the most history about you. Connectors are the pipes into the apps you already live in, and they let your AI act inside them.

- **ChatGPT:** Settings → Connectors → add Gmail and Google Calendar. Gmail and Calendar need Plus or Pro. On the free plan, skip this step; the interview does the work.
- **Claude:** Settings → Connectors → add Gmail, Google Calendar and Google Drive. Every plan, free included. Web, desktop and mobile.
- **Gemini:** Settings & help → Connected Apps → turn on Google Workspace. One switch covers Gmail, Calendar and Drive. Every plan, free included; needs Keep Activity on.
- **Terminal:** Ask it: "Give yourself persistent access to my email and calendar for future sessions, and tell me what you set up." It names the connector or CLI it needs; approve it once. Check what it can now read before you move on.

Worth a look, in both directories today: Gmail · Google Calendar · Google Drive · Slack · Notion · Stripe · Shopify · Canva · HubSpot · Klaviyo · Dropbox · Zapier. Creator tools: vidIQ and Ubersuggest (ChatGPT), Metricool and Buffer (Claude). Health: ChatGPT's Health app reads Apple Health, US only, free plans included.

One habit from the talk: separate brains for separate jobs. Each gets only the connectors that job needs.

### 4. Put it to work

**The clock.** A schedule is a time ("every Monday at 7"). A trigger is an event ("when a brand email lands"). Eight tasks to put on the clock, copy them as written. Turn on the weekly review first; it needs nothing connected:

- **Weekly review (sends to you):** "Every Sunday at 6 p.m., look at my calendar and inbox from the week and what we talked about this week, whichever of those you can reach. Ask me three questions about my week, then tell me the one thing to fix." (needs: nothing; better with Calendar and Gmail)
- **Watch a topic (sends to you):** "Every Friday, search for what is new this week on [a competitor, a market, a condition] and send me only what changed and why it matters to me." (needs: nothing; fill in the bracket)
- **Morning brief (sends to you):** "Every morning at 6, send me today's calendar, yesterday's key numbers, and the one thing that matters most." (needs: Calendar; a payments connector for the numbers)
- **Meeting prep (sends to you):** "Every morning at 7, for each meeting on my calendar today, send me who I am meeting, the last thing we exchanged, and three questions worth asking." (needs: Calendar and Gmail)
- **Weekly numbers (sends to you):** "Every Monday at 7 a.m., pull last week's revenue, new customers and cancellations, tell me what changed, and give me one likely reason." (needs: a payments connector (Stripe, Shopify))
- **Follow-up nudge (drafts only):** "Every weekday at 4 p.m., find emails I sent that have had no reply for three days and draft a short follow-up in my voice. Don't send." (needs: Gmail)
- **Brand-deal sweep (drafts only):** "Every Monday at 7 a.m., find new sponsorship or brand emails, list each one with its stage (inquiry, negotiating, signed, delivered, paid), and draft replies. Don't send." (needs: Gmail)
- **Friday repurpose (drafts only):** "Every Friday at 3 p.m., take what I wrote and said this week and draft next week's post or newsletter in my voice. Don't publish." (needs: nothing; better with Drive or Notion)

Where: ChatGPT scheduled tasks (free plans get 3, on a flexible window; paid plans pick exact times, and only paid plans get triggers), Claude scheduled tasks, Claude Code routines, Gemini scheduled actions (paid plans first). One ChatGPT rule: a task made inside a Project can't read the Project's files, so keep the task's own words self-contained, or have it send you back into the Project.

**Permissions.** You decide, tool by tool and job by job, whether it drafts or sends.

| Rule | What it means | Examples |
| --- | --- | --- |
| Draft | Drafts only, you send | Money, contracts, new brands, public posts under your name |
| Ask | Asks first, one tap on your phone | Replies in existing threads, follow-ups, approved posts |
| Send | Sends on its own, logged | Anything sent only to you; proven templates |
| Never | Blocked | Spending, deleting, passwords, security settings |

Where to set it: Claude uses Always allow / Needs approval / Blocked per connector. ChatGPT uses Always ask / Allow read / Allow low-risk / Allow all per app. New tool, new rookie: it starts on Draft and earns its way up.

**The review.** A system that runs on its own drifts on its own. Put a 45-minute "OS review" on your calendar, monthly, and open it with: "Read my file and last month's notes. What drifted? What did I fail at twice? Is the goal wrong, or is it me?"

## Memory, in one paragraph

Turn on the app's own memory and stop there: in ChatGPT it's Reference chat history (Settings › Personalization › Memory); in Claude it's Search and reference chats (Settings › Memory, paid plans). Be skeptical of anything else sold as "memory." For a project or a business, the real memory is documentation with history: a GitHub repo and good docs, with a pipeline that turns every solved problem into a doc your AI reads next time. Will's name for that pipeline is compound engineering.

## A second brain with AI

A second brain used to mean notes you filed and searched yourself. With AI, the brain does the reading: it opens your file, checks the app's memory and your connected Gmail, calendar and files, then answers like someone who already knows your goals. Steps 1 to 3 build it. Step 4 puts it to work, with a schedule, permissions you set job by job, and a monthly review.

## Go further

Everything above works in the web app on your phone. When you want your AI to take actions on your computer, the terminal is the next level (Claude Code, Codex). Above that sit the always-on assistants (Hermes, OpenClaw, Meta's Muse, Grok's bot); Will builds and hosts his own instead. The tools he mentioned are open source on the House of Vibe marketplace, as Claude Code plugins in [hov-stack](https://houseofvibe.ai/hov-stack), his Claude Code setup:

- [memory-dream](https://github.com/StartupBros-com/hov-stack/tree/main/plugins/memory-dream): keeps the memory files from rotting
- [papercut](https://github.com/StartupBros-com/hov-stack/tree/main/plugins/papercut): lets agents file their own frictions
- [sign-it](https://github.com/StartupBros-com/hov-stack/tree/main/plugins/sign-it): "sign that" on a contract PDF
- [harness-vet](https://github.com/StartupBros-com/hov-stack/tree/main/plugins/harness-vet): vets a skill or plugin before you adopt it
- [rent-check](https://github.com/StartupBros-com/hov-stack/tree/main/plugins/rent-check): every always-loaded line has to pay rent

## Bonus: Will's own folder

The talk shows one file on screen. This is the folder it lives in, as it is on his machine, minus the finance and health exports.

```text
will-os/
├── README.md          quick reference
├── CONSTITUTION.md    principles and boundaries
├── CLAUDE.md          how the AI behaves in this folder
├── theme.md           the year's theme, history below it
├── rocks.md           the quarter's rocks, history below it
├── inbox.md           processed items
├── outsource.md       time leaks to hand off
└── monthly/           one review per month
```

How often it runs:

- **Daily:** capture to the to-do app, work from paper (5 min)
- **Monthly:** the review (45 min)
- **Quarterly:** set the rocks (30 to 45 min)
- **Yearly:** set the theme, then the first quarter's rocks (90 min, then 45)

How the AI is told to talk to him, from the folder's CLAUDE.md: Direct. No education — Will knows the principles. Challenge when appropriate. One question at a time. No emojis. No cheerleading. If he pushes back on a flag, note it and move on. He has final authority.

What it flags, in one sentence each:

- **Productivity theater:** "What rock does this serve?"
- **Scope creep:** "Which rock does this replace?"
- **Outsourceable work:** "Log it?" (costed at your hourly number)
- **System tinkering:** "Is this necessary or procrastination?"

His whole CONSTITUTION.md, the file his AI reads before it does anything for him: https://houseofvibe.ai/personal-os/constitution.md (raw Markdown; paste it into your Project as a starting point, then make it yours; two edits: his hourly figure reads [your hourly number], and the dated footer is dropped).

Before the interview prompt, one more thing worth running: the Reflection Engine, https://github.com/StartupBros-com/reflection-engine, a copy-and-paste prompt that makes the AI you already talk to tell you what it has noticed about you. Let it tell you who you've been; the interview writes who you decide to be.

The slides, all 44: view them in the browser at https://houseofvibe.ai/personal-os/deck or download the PDF (3.6 MB) at https://houseofvibe.ai/personal-os/deck.pdf.

What the room said: the final tally from the session is at https://houseofvibe.ai/personal-os/poll/results.

## Will

LinkedIn: https://www.linkedin.com/in/willmitchell · X: https://x.com/startupbros · willmitchell.com: https://willmitchell.com

## House of Vibe

If you want to build this with a room of people doing it every month, that's House of Vibe. StartupBros' membership for non-technical people building with AI: a guided project library, a live workshop every month with every session recorded, a private community, and every future track included. It is one purchase with lifetime access. https://houseofvibe.ai/?utm_source=personal-os&utm_medium=handout&utm_campaign=mic-2026

Machine-readable siblings: https://houseofvibe.ai/personal-os.md (this file) and https://houseofvibe.ai/llms.txt.
