Sasha is the person who runs the day around you — reads your mail first, preps you for every meeting, keeps track of what's open — a personal AI that's genuinely yours, one anyone can stand up.
One copy-paste prompt turns Claude Code into a personal operating system, wired to your Google (Gmail + Calendar) or Microsoft 365. It runs your week — triage, meeting prep, status updates — and gets sharper every time you correct it. Five questions and one paste — most people read their first brief in about ten minutes.
no code required · read-only by default · nothing sent without your sign-off
New here? The Field Guide → shows how to get the most from every command — plain-language up top, power-user depth below.
No code, no terminal. Install one app, connect your mail & calendar, switch on Auto, paste the prompt — then Sasha asks five quick questions and builds around you. Most people read their first brief in about ten minutes.
Download the Claude desktop app from claude.com/download — a normal app installer for Mac (.dmg) or Windows (.exe). Open it and sign in with your Claude account. You never need a terminal or any code. The app runs on a paid Claude plan — Pro is about $20/month. This setup is free; the AI that powers it isn’t.
Windows detour (one time, a couple of extra minutes): the first time you open it, the app may ask you to install Git for Windows — click Install, accept the defaults, let it finish, then reopen the Claude app. That’s the whole detour. (Mac has nothing extra.)
Prefer the command line? The Claude Code CLI at claude.com/claude-code is the very same engine — every step below works there too.
On your Claude account at claude.ai → Settings → Connectors, turn on Google (Gmail + Calendar) or Microsoft 365 and sign in to authorize it. Your AI OS works read-and-draft: it reads your mail and calendar, drafts locally, and anything outward waits for your OK — that per-action approval is the guarantee on either provider. On Microsoft 365 the connector itself is read-only on top, which takes nothing away: Microsoft 365 is every bit as capable as Google here. The difference is access, not capability.
Pick the one you can actually connect. Google is quickest — most people, personal Gmail included, can connect it themselves. Microsoft 365 needs a work account your organization has enabled — a personal @outlook.com or @hotmail.com address can’t connect at all.
Most people can approve it themselves; some tenants still need a one-time admin consent — so confirm you can connect before you invest time. If the connection ever needs a fresh sign-in, Claude tells you and reconnects the moment you re-authorize. The setup prompt can walk you through it.
In the app, click the Code tab, then create or pick a folder to be Sasha’s home — e.g. a new folder called ai-os. That folder is where your memory and notes live, on your own machine.
The one step people miss. The mode control sits in the top-left corner, just above where you type — switch it to Auto and accept the one-time confirmation dialog. (If Auto isn’t offered there, enable it once in the app’s Settings.)
Auto means the build runs in one pass instead of stopping at every file — and the guardrail that matters doesn’t move: your mail and calendar stay read-and-draft, and anything outward waits for your per-action OK (on Microsoft 365, the connector itself can’t write at all). Near the end it asks permission to schedule the quiet data refreshes that keep your morning brief instant; that’s expected — say yes.
Copy it below, paste the whole thing, and send.
Sasha introduces herself and asks five questions: what fills your week, which account to wire up, what to take off your plate first, what she should never do without asking, and how you like things. Say “skip” to any of them — she starts with sensible defaults and learns as you work. Then she builds everything shaped to you and explains what she made.
The moment the build finishes, it hands you your actual day — today's meetings, what's waiting on you, deadlines this week — proof it's wired to your real mail and calendar. That's the payoff, and you didn't have to do a thing. From your next session on, type /brief any morning to get it again for that day; your other commands (/triage, /prep, /recall) work the same way — close this folder and reopen it (or start a new chat) once, and they're all ready.
Loading…
❯ the one habit that makes it pay off: run /improve whenever you correct it — that's the loop that makes it more you.
Already set up? Pasting the prompt again upgrades in place — it detects your install and only adds what's new, leaving your memory, notes, and customizations untouched. Run /doctor anytime and it'll check whether you're on the current build.
Picture a normal Monday: back-to-back calls, a weekend's worth of unread mail, a status update due Friday, and the thing you promised someone you'd track. Hand that recurring overhead to an assistant that already knows your work — and watch the dots go green.
The prompt interviews you about your role and your work first, then builds all three around what you actually do — not a generic template.
Claude Code remembers across sessions — this keeps that memory honest: it re-checks what's gone stale, clears duplicates, resolves contradictions, and links related facts, so what it knows about your projects and people stays trustworthy instead of drifting.
An Obsidian-style vault for projects, the people you work with, decisions, and reusable templates — status update, meeting notes, weekly review, 1:1 agenda.
Slash-commands you invoke by name — /brief, /prep, /triage, /weekly, /ingest — each wired to your mail & calendar.
Type a slash-command and it does the work — reading from a fast local cache of your mail & calendar, drafting locally, and sending nothing without your explicit OK.
Plus /memory-prune to keep your memory healthy, and /skill-center — a meta-command to build and refine new commands as your needs grow.
Want to go deeper? The Field Guide → walks through how to get the most from each command — plain-language up top, power-user depth below.
Designed to be trustworthy from the first line. Three controls keep it fast, in your hands, and private.
Your mail, calendar, and notes live only on your own machine — a hard "never goes outward" rule. The one thing that can leave is a backup you ask for: encrypted, into your own cloud, so your system survives a lost laptop without ever being exposed.
Every command reads and drafts; nothing is sent, posted, accepted, or shared until you approve it. Approving one action is never standing approval for the next.
It caches your calendar and mail and refreshes only when something changes — instant to use, and gentle enough on your mail and calendar service that it won't trip rate limits.
Most "AI second brain" tools are hosted apps: your notes live in their cloud, on their terms, for as long as you keep paying. An AI OS is the opposite — a free template you copy, run on tools you already use, and reshape to fit you. Here's the honest comparison.
Every tool below is a paid SaaS account you rent a seat in. This is a free copy-paste template — you can read it, reshape it, fork it, and pass it to a friend. There's no vendor to leave because there's no vendor.
Your notes and memory are plain text files on your own machine — the store is the data. You can open, search, edit, and version them directly, with no export step and nothing trapped in a proprietary cloud database.1
| AI OS | Saner.ai | Mem | Tana | Notion AI | ChatGPT / Claude memory | |
|---|---|---|---|---|---|---|
| Where it lives | Notes & memory: local files on your machine. Model & connectors: cloud1 | Cloud app | Cloud app | Cloud only (no offline) | Cloud app | Cloud, tied to your account |
| Your data's form | Plain markdown — the store is the data; edit & search in place, no export step | Proprietary cloud; export available | Proprietary cloud; markdown export (clunky) | Proprietary nodes; JSON/MD export, but lossy | Proprietary blocks; full export | Cloud summary you can view/edit + a data export — not files |
| Own & fork it | Free template — fork, reshape, pass it on | Rented seat | Rented seat | Rented seat | Rented seat | Rented seat |
| What it does on its own | Reads & drafts; every outward action waits for your per-action OK — on by default, no mode to switch on (on M365 the connector is read-only outright) | Drafts first; doesn't auto-send | Notes & drafting assistant | Knowledge workspace | Agents act autonomously by default; approval (Plan Mode) is opt-in | — (it's memory, not actions) |
| What it costs | Free template — runs on the Claude plan you already pay for | Paid subscription | Paid subscription (free tier) | Paid subscription (free tier) | Paid seats + metered agent credits ($10 / 1,000, no rollover) | Bundled in your ChatGPT / Claude subscription |
| Keeps memory healthy | A maintenance loop you can run & audit: re-checks stale facts, clears duplicates, resolves contradictions, links related notes | Auto-tags / organizes | Auto-links related notes | User-defined supertags | Workspace knowledge (not a personal memory) | Background synthesis + an editable summary; claims to reduce contradictions — but an opaque cloud summary, not files |
1 Honest scope: your notes and memory are plain files on your own machine; the AI model and the mail/calendar connectors are cloud services you already use. An AI OS doesn't run offline, and we don't claim it does.
Every tool here lets you export your data — the real difference is that theirs lives in a cloud database you export from, while yours is already plain files you own. Competitor details are as-of July 2026; hosted-app features and pricing change often — check the source before relying on any cell.
Setup already scheduled the quiet data refreshes that keep your commands instant — this is different: scheduling the brief itself, so it’s written and waiting when you sit down. When your /brief and /triage are part of your morning, the next step is to stop typing them. Claude’s desktop app can run a command for you on a schedule — a 7am brief waiting when you sit down — and because it runs on your machine in Ask mode, it still drafts and waits: anything it would send stalls for your OK, exactly like when you’re driving it yourself. Nothing is sent while you’re away.
Read this part — scheduling has real edges:
· It only runs while the app is open and your machine is awake. A sleeping laptop silently skips the run (it catches up on wake, maybe at an odd hour) — pick a time your computer is on.
· A green “done” doesn’t prove the brief worked. Open the run and read it.
· Every unattended run uses your Claude plan like a normal session — a daily brief is a recurring draw that can crowd a ~$20 Pro plan, so start with one task a few days a week and watch your usage.
Paste the prompt below and it walks you through setup and hands you the exact text to schedule.
Loading…
❯ the rule that keeps it safe: schedule it in Ask mode, and on the first “Run now” never click “always allow” on anything that sends — that’s what keeps it drafting instead of acting.
Your AI OS runs your day. The next step is to hire a domain manager — a standing chief-of-staff for one part of your life that keeps its own strategy, does its own research, stages real work for you to approve, and gets sharper every time you correct it. Think of it the way you’d staff a small team: you’re the boss, you hire one manager at a time, and nothing they draft goes anywhere until you say so. It’s a bench of focused workspaces you open when you want them — not an autonomous army of bots.
Learns how you write and drafts LinkedIn posts in your voice; keeps a living strategy for where your career is headed. Read the full role, then open the career prompt below to hire it.
Keeps a strategy, backlog, and decision log for a project you own — and drafts the weekly status you actually send. (It’s the same shape that runs this product.) Read the full role, then open the projects prompt below to hire it.
Turns statements you paste in into a budget and a plain-language read on your money. It drafts and explains — it never moves a cent.
Plans what’s worth learning across everything you’re trying to master — a prioritized, prerequisite-ordered roadmap — then hands each topic to your tutor (/teach). Read the full role, then open the learning prompt below to hire it.
Career, Projects and Learning are the three you can paste today. Finances is the same pattern, on the way — and because it’s all plain files you own, an expert can stand one up now by reusing the shape.
Read the full role → what it does, when to reach for it, the honest limits.
The career manager learns how you write from a few things you’ve written and drafts LinkedIn posts that sound like you, not generic AI. It won’t nail your voice from three samples — no tool does — so it gets it roughly right, shows you what it learned, and sharpens every time you edit a draft. You’re always the judge, and you do the posting — nothing is ever published for you.
Loading…
❯ the loop that makes it you: every time you reword one of its drafts, run /career voice learn — it remembers the difference.
Read the full role → what it does, when to reach for it, the honest limits.
The projects manager — the same STRATEGY / BACKLOG / DECISIONS shape that runs this very product. Paste it, answer a short interview about one project you own (its goal, who reads the status, how often you report), and it drafts the weekly status update you actually send — pulled from the plan, not written from scratch. You’re always the judge, and you do the sending — nothing is ever sent for you.
Loading…
❯ once it’s running, type /projects update anytime and it drafts your next status from where things actually stand.
Read the full role → what it does, when to reach for it, the honest limits.
The learning manager — the same STRATEGY / BACKLOG / DECISIONS shape, pointed at everything you’re trying to learn. Answer a short interview about your goals and it drafts a prioritized, prerequisite-ordered roadmap — what to learn, in what order, and where to start — then hands each topic to /teach, the tutor that already ships with your AI OS. It plans; /teach teaches. No connector, nothing to send — just a standing strategist that stops you learning at random.
Loading…
❯ once it’s running, type /learning review anytime and it tells you which topics have gone cold and are worth revisiting.