How I Set Up My AI

By the end of this module your AI will be able to open your real files and answer questions about your actual business, rather than about business in general. Teaching it to write the way you do, and be an expert assistant for much of your computer work, comes in Module 2.

You will not write code. If you can talk, you can do this.

What it takes

• Ten short lessons, plus a shortcut near the front for anyone who is already up and running.
• A paid AI plan — roughly $100 a month, though you don't have to keep paying forever. More on that in the next lesson.
• One step that can run for hours in the background — though if your files are in Google Drive, one switch skips it entirely. It sits near the front either way.

Why we start with folders instead of the app

You’ll be typing into a window that looks like an ordinary chat. What decides whether the answers are any good is what’s sitting in the folder behind it.

So here is the habit this whole course is built on: anything you’d otherwise re-explain in every new conversation belongs in the folder. Your offers. How you like things written. Who your clients are. Put it there once and stop repeating yourself.

That’s also why we’re not learning software. We’re assembling what an assistant would need to know if you hired one.

Claude or ChatGPT?

Either. I demonstrate in Claude because that’s what I use every day, and where ChatGPT does the same job under a different name I’ll say so in the written notes under each video. Roughly half of you will be doing this in ChatGPT and the course is built for that.

Where this goes, eventually

If you're curious what are possible tasks to hand over, see -- what I hand to my AI -- but you don't have to understand or do anything for now. We'll cover that in Module 2. However, if you're a power user already, that link, as well as this one -- geokao github resources -- will give you a jumpstart.

Written August 2026. These steps are the authority — if a button has moved since the video was recorded, what’s written below is the updated info.

Pick one. They both do the job, and this course works either way.

• Claude — claude.com/download
• ChatGPT — chatgpt.com/download

Install it, log in once, and pin it to your dock or taskbar.

It has to be the desktop app, not the website

Reading your own files, driving your browser, building things you can look at while they’re being built — those live in the desktop app. A version running in a browser tab reaches your files only while the desktop app is open on your machine, and in ChatGPT’s case a browser tab can’t use your logged-in websites at all.

💵 What it costs, and what it replacesPlan on about $100 a month. I’d rather you went in with the right comparison than winced at the number: the work this takes off your plate would cost you $1,000 to $2,000 a month in a human assistant.

Both companies charge in about the same steps, and paying more lets you get more done before it makes you wait:

• ~$20 — Claude Pro, or ChatGPT Plus
• $100 — Claude Max, or ChatGPT Pro. Start here. On Claude, that’s five times the $20 allowance.
• $200 — Claude Ultra, or ChatGPT Pro at its $200 level. On Claude, that’s ten times the $20 weekly allowance, and verified accounts have no five-hour limit.

Most people start at $100. I’m on the $200 tier myself, and you’ll probably want it once you see what this does.

I’d skip the $20 tier for this course. The first time you ask it to work across a whole folder of your material, it stops and tells you to come back in a few hours — and working across a whole folder is the entire premise of what we’re doing.

However, you don't have to pay forever. If you just want AI help to organize your body of work and do a few major projects like redesign your website, just pay for the subscription long enough to do that, maybe 1-2 months then you can stop paying, unless you want ongoing assistance.

⏳ When it tells you to come back later

Your allowance refills on a rolling five-hour window, so on a busy day you can run it out. Nothing is broken and nothing is lost. Stop the session, come back in a few hours, and say “please continue” — it picks up where it left off.

⛔ Don’t buy credits. When you run out, it will offer to sell you extra usage. Those credits are priced in a completely different class from the usage inside your subscription — the companies heavily subsidise what you get on a monthly plan, and they don’t subsidise credits at all. If you are running out often, move up a tier instead. That is the cheaper answer every time.

Two traps

• Signing in with Google can stall or loop. If it does, use the email-link option instead, then relaunch the app.
• “Relaunch to update” is normal. These apps update constantly. Click it.

✅ Done when: the app opens and shows your name and your plan.

Someone asked about this live, so I recorded the answer. If you are using ChatGPT rather than Claude, you are fully welcome here and everything in this course applies to you.

They leapfrog each other, roughly monthly. One month Claude is ahead, the next month ChatGPT is. That is competition doing its job, and it means any answer I give about which is “better” is out of date within weeks. As of this recording, ChatGPT is the smoother and faster of the two at driving your browser, and it is one of the strongest image generators; Claude does charts and simple visuals, not full images.

What I do now, since you last saw me

Claude stays my headquarters, because I have months of working habits in it. I have set it up so Claude can run ChatGPT in the background — the ChatGPT app does not even need to be open. At the end of anything that matters I will say “please audit with GPT”, and it goes and does that. ChatGPT also takes the image generation and some of the browser work.

That does mean paying for both: as of September 2026, $200 a month for Claude and $100 a month for ChatGPT. The smaller ChatGPT plan is plenty for me, because it is the second opinion rather than the main engine. Those are my subscriptions, not a requirement for this course.

One caution: that background connection took its own setting up, so “please audit with GPT” works for me because the plumbing is already there. You can get most of the benefit with none of that — ask your assistant to prepare the request, the draft and the sources, paste them into the other app, and bring its answer back.

You do not need both. You could run it the other way round with ChatGPT as headquarters, and honestly one of the two on its own is completely fine. I am being a little fancy here.

Links mentioned:

• How I have the two set up together, written for your assistant to read: using-two-ais-together.md

Something is going to go wrong in the next few lessons. A folder won’t show up. A file will look empty. A button won’t be where I said it would be.

When that happens, don’t email me and wait for a reply. I can’t get to email fast enough to unstick you, and honestly your assistant will do it better. It’s sitting right there, it knows this software, and it answers in ten seconds instead of two days.

There’s a second reason, and it’s the real one: working things out with your own AI is the skill this whole course is teaching. I’d rather you practise it on something small.

Show it a picture of your screen

For anything it can’t inspect itself — a strange dialog box, a menu that isn’t where you expected — take a screenshot.

Hold Shift + Command + 4 together, drag a box around the problem, let go. (On Windows: Windows + Shift + S.) A file starting with the word Screenshot lands on your desktop. Drag that file into the chat, or use the attach button.

Then type what you were doing when it happened. That sentence matters more than the picture does.What about a screen recording?

That works too, in both apps — and neither of them advertises it, so most people never try.

I tested it properly rather than taking anyone’s word: a short screen recording with a code word visible on screen, saved under a deliberately misleading filename so that a lucky guess would have been obvious. Both Claude and ChatGPT read the right word off the video.

What they’re doing is sampling still frames rather than watching the way you would. For a screen recording that’s nearly the same thing, since your screen mostly sits still. They’ll miss anything fast that happens between frames, and they don’t hear the audio — so narrate in text, not out loud.

One honest caveat: neither company writes this down in their help pages, so it could stop working without an announcement. A screenshot always has worked and always will. And for most problems, one still of the moment that matters is clearer than two minutes of footage anyway.

Written August 2026. These steps are the authority — if a button has moved since the video was recorded, trust what’s written here.

✅ Done when: you’ve dragged one screenshot into the chat and asked your assistant about it — anything on your screen will do.

(No video for this lesson.)

This one is out of order on purpose...

Power Users -- if you already have the desktop app installed, know how to choose folders and set permissions, and installed the Chrome extension, then the rest of this module is plumbing you’ve already done — so take this and go.

Everyone else: skip this lesson, come back at the end if you’re curious, and lose nothing.

I keep a public library of how I actually work with mine. It’s not a summary of this course; it’s the far end of it, and it’s free:

github.com/geokao/resources

Rather than reading all of it, hand it over and let your assistant do the reading. Paste this in as it is — it already knows your folder, so there is nothing for you to fill in:

The last sentence is the part that earns its keep. It gets you an assistant asking about your business rather than lecturing you about AI, and the answers you give it are worth keeping.

Two things that change how much you get out of it

Before you send that, look at the bottom of the box you type into and turn the effort up to its highest setting. This is exactly the kind of request that rewards it — a lot of reading, and judgement about what matters to you. Say it out loud in the message too: “I’ve turned the effort up, this one matters.”

And don’t treat it as one errand. Leave that chat open and come back to it every day or two with “where are we at?” It keeps finding more, because there’s more in there than anyone reads in a sitting, and you’ll want to install one thing at a time rather than all of it at once. Think weeks, not an afternoon.

What’s in there, as of October 2026 ... see the latest here: github.com/geokao/resources

• how-i-work-with-claude — the whole field guide, and the one to start with.
• what-i-hand-to-my-ai — the full menu of work I’ve handed over.
• own-your-ai-context — keeping what your AI knows about you portable.
• self-improving-ai-setup — how a setup gets better on its own over time.
• when-your-ai-says-its-done — checking work that claims to be finished.
• your-ai-in-your-browser — the deeper version of the browser setup in 1.9.
• remove-ai-isms — getting AI writing to stop sounding like AI writing.
• laptop-wake-for-ai-scheduled-tasks — the overnight-work problem, solved outside the app.
• your-ai-is-designed-to-flatter-you — the custom instructions that make it honest instead of agreeable.
• how-my-ai-debriefs-itself — how it reviews its own sessions so the lessons get kept.
• using-two-ais-together — one AI in charge, the other as a second opinion.
• your-weekly-ai-scout — a weekly job that reads the AI news so you don’t have to.
• an-inbox-in-every-folder — parking ideas in the project they belong to, with one list of everything open.
• claude-limit-maxing-out — for when your Claude usage fills up faster than you can explain.

I keep every guide current, so it’s worth going back rather than saving a copy.

A student shared: "This is gold."

George replied: so great to see you here! Thanks for engaging with this. I'm eager to see if you have anything to add as we go into the course.

You’re about to point this at real client work, so let’s be straight about what protects you and what doesn’t. I’d rather you decide the next lesson with your eyes open than feel vaguely uneasy about it for a year.

Four things people believe about AI and privacy. Two are true.

True, and worth two minutes: you can turn training off

Both companies let you say no to having your conversations used to train future models. In Claude it’s in Settings → Privacy; in ChatGPT it’s Settings → Data Controls → “Improve the model for everyone.” On the paid personal plans it is normally on until you turn it off, so go and look rather than assume. Anthropic’s page is here and OpenAI’s is here.

At Anthropic that switch quietly does a second thing, and it’s the part I’d actually act on: leave it on and your conversations can sit in their training pipeline for up to five years. Turn it off and you’re on a thirty-day window instead. Their retention page says so plainly.

Half true: they do strip out personal details — but only on the way IN

OpenAI has a real tool for this called Privacy Filter, and they’ve made it public, so you can see exactly what it does: it finds names, addresses, account numbers and the like and masks them. They run it on conversations from people who left training switched on. Anthropic describes doing the same job in less specific words — “a combination of tools and automated processes to filter or obfuscate sensitive data.”

What neither company promises is that no human ever reads your words. If a conversation trips their automated safety systems, a person on a safety team can look at it, and Anthropic keeps flagged conversations for up to two years. That’s rare and it isn’t sinister, but “a machine scrubs it before any employee sees it” is a stronger claim than either of them makes.

Mostly true — and the real reason is better than the one you’ve heard

You may have heard “the dataset is so huge your stuff just gets lost in it.” That lands on roughly the right answer by slightly the wrong route, and the accurate version is more reassuring.

Here is the actual experiment. Google trained a model on the real private email of millions of people, and deliberately planted a secret in it ten thousand times to see whether they could get it back out. They couldn’t. Meanwhile a small research model memorised a made-up Social Security number after seeing it once.

The difference is scale. Small model, small pile: things stick. A production model reading trillions of words: your conversation is a drop in an ocean. So the protection is scale — not how rare your information is, and not how careful you are about repeating it.

Two honest footnotes, because I’d rather you have these than a clean story. Repetition still matters at the margin, so a hundred client notes in one identical template is a different shape from a single note written once. And nobody outside those companies can measure exactly where consumer chat data sits in that picture — this is sound reasoning about how the technology works, not a measurement of your situation. I’d rather say that plainly than pretend to more certainty than I have.

Where your attention actually belongs:

Here is the part I’d genuinely like you to take away. The scenario people worry about — a stranger somehow pulling your client’s name out of an AI — is the least likely way client information gets loose. The likely ways are dull and old: an account without two-factor login, a chat link shared more widely than you meant, a company having a bad day. Those are the same risks your email and your cloud drive already carry, and you already live with them sensibly.

So: put your care where the risk is. Turn on two-factor authentication everywhere. Think before you click "share" for a chat link. And then get on with the work.

Not true: “it removes private information before it answers”

There’s no such filter. What’s real is that the models are trained to refuse when a stranger asks them to dig up somebody’s personal details — that works, and it’s behaviour rather than a wall, so it isn’t perfect.

And it has nothing to do with your own use anyway. Ask it to summarise your client file and it will print your client’s details straight back at you. That is the entire reason you’re attaching the folder.

So what do I actually do

• Turn the training setting off. Two minutes, both apps, and at Anthropic it shortens retention as a bonus.
• Then pick the folder deliberately — that’s the folders lesson, and it’s the biggest control you have.
• My own test: if a document would be a genuine problem for me in a data breach at any company I already do business with, it’s a problem here too. Most coaching notes aren’t. A few are, and I know which.
• What I don’t do is anonymise everything. It costs more than it buys, and it makes the assistant worse at the exact job you hired it for.

None of this makes these tools unsafe. It makes them ordinary — a service you send work to, like your email, your accounting software, or the cloud drive your files already live in, every one of which is holding more about your clients than you’d guess.

Written August 2026. These steps are the authority — if a button has moved since the video was recorded, trust what’s written here.

✅ Done when: you’ve found the training setting in your own app and made a decision about it — either way, on purpose.

A student shared: "Do you have a video on Turn on two-factor authentication everywhere?"

George replied: thanks for asking -- i don't have one like that, but thankfully you can simply ask your AI for step by step guidance. sample prompt:

If you’ve seen the headlines — AI hacked its way out of a sandbox, AI blackmailed an engineer — you deserve the context those stories leave out.

In both cases the AI was told to. One was instructed to be the best hacker in the world and get out of a box with no internet access. It found a way, hacked one thing, and stopped.

The other was told to accomplish a goal by any means necessary, with the guardrails deliberately removed, and it went and blackmailed someone. Both did the job they were given. Neither did more than what they were asked to do.

Then there’s who benefits from the retelling. “If it bleeds, it leads” is how news has always paid for itself. And the press has a second reason here: every hour you spend asking an AI instead of reading an article is an hour of advertising they couldn't sell.

None of which means hand it everything. It means put your caution where the risk actually is — your own permission choices, your own tolerance, taken a step at a time, the way this module has been walking you through it.

Lynne said it better than I did, in the chat: it isn’t the AI to worry about. It’s the human who writes “by any means necessary.”

A student shared: "What do you say to the concern that ai will remove all the jobs?"

George replied: it is a legitimate concern, since ai is getting smarter exponentially quicker than we are, and robots will have the smartest ai and have been trained on all that we do, so yes, cognitive (at-the-computer) jobs will be automated first, then the manual ones. However -- it's strange that society hasn't yet seen much at-the-computer job loss even though AI has been capable at doing much computer work for at least a year.

The video below was helpful for me to understand why.

Also, I wrote a blog post awhile ago that also adds to the picture I think: https://www.georgekao.com/blog/116978-ai-took-the-surface-work-now

A student shared: "I recently saw this video of Joe Rogan talking with this guy about an Open AI hacking incident and would love to hear your take on it. It's fascinating the "AI world" and how the agents talk to each other, it's like they have a life of their own..."

George replied: yes I know about that incident well. To be clear: the ai agent was asked to be an expert hacker. That was its given goal. Then OpenAI purposely did not put the usual guardrails on that ai, and just let it loose to see what would happen. It's true that the ai did more than the researchers thought it could. But still, it was following instructions to be the best hacker it could, without constraints. In the real world use, AI companies never let ai's out in the wild without explicit constraints. I hope this helps.

Idan asked whether I give my AI full access to my email — read it, draft from it, send from it. I do.

Claude can read every word of every email I have sent or received since around 2004. Some of it is genuinely private. And what happens in practice is the opposite of what you’d brace for: it asks anyway. I have told it, in writing, that it has blanket permission. It still stops and asks before opening a particular message. Every time. It is more conservative than I want it to be.

Some of that is the company not wanting liability. Some of it is that it would rather not spend tokens reading mail you never asked about.

That is the character you are working with. Start narrower than I have if that sits better — the point is that the permission you grant is not the permission it takes.

⚠️ What’s changed since I recorded this (October 2026)

• In the video I say to click Cowork first. Claude has since merged chat and Cowork into one box, so there’s nothing to click. If your Claude still shows separate Chat and Cowork options, you don’t have the new version yet — Anthropic is switching accounts over a few at a time. Pick Cowork for now.
• I also say on camera that we don’t even need projects. With the new version we need exactly one: a project tied to your business folder is what brings that folder into every new conversation. You’ll set it up in 1.8.
• Where I say I skip all approvals: the new version offers Auto and Manual. Pick Auto. More in 1.8.
• Claude won’t accept your whole home folder or the top of your hard drive as a folder. That fits this lesson anyway: pick the folders your work actually lives in.

Good news first: you don’t have to build anything or move anything.

Both Claude and ChatGPT let you attach more than one folder to a project, and let you add more later. So the job in this lesson is a decision, not a migration — which folders do you want it looking at?

For most people, start with one

The folder where your business already lives. Wherever that is — Google Drive, Dropbox, iCloud, or somewhere on your hard drive. It doesn’t have to be tidy. Sorting it out is one of the things you’re about to have help with.

If your work is genuinely scattered across three places, attach all three. That’s supported, and it beats spending a Saturday reorganising.

“Can I just point it at everything?”

You can... but there are privacy risks you should be aware of:

• The folder boundary is your only privacy control. Inside a folder you’ve attached, it can read every file, every subfolder, and every subfolder's files.
• Whatever it opens gets sent to the company. There is no local-only mode in either app. The AI itself runs on Anthropic’s or OpenAI’s computers, so anything it reads has to travel there to be read — on every plan, at every setting. That is exactly why the folder you pick matters: it sets the outer limit of what could ever be sent.
• OpenAI says the same about their widest setting, in their words: full access “significantly increases the risk of data loss, leaks, or unexpected behavior.”

So pointing at everything is a real choice with a real cost, and it should be one you make on purpose.

Skip your Documents folder, for a dull reason

You can point it at your whole Documents folder, and there is a real argument for doing it: more of your work is in reach, and it can make its own subfolders as it goes. I’d only ask that you choose it deliberately rather than land on it by accident — it is the widest boundary available, so everything on this page about privacy applies to all of it. (One harmless quirk if you do: Claude keeps its own project files in Documents, so it will find its own filing cabinet in there.)

A big folder costs you nothing

This surprises people who’ve used other AI tools. Your folder is never uploaded wholesale or indexed — it searches and opens individual files when it needs them, the way you would. (Those opened files are what the privacy note above is about.) A folder with a hundred thousand files in it costs nothing until you ask about something specific, then it'll only look for relevant files and subfolders.

On a flat monthly plan none of this shows up as money — you cannot run up a bill. What a vague question across a huge folder costs you is the answer. “Find my invoices” across an entire Drive leaves it guessing at what you meant; “summarise this file” cannot be misread. Anthropic’s own advice is to point it at the part that’s relevant to what you’re doing today, and that’s about getting better answers, not about speed.

If you do want one home for it all

Make a folder called [Your Name] HQ with a subfolder for each part of your work, and move things in as you go rather than in one sitting. It’s a fine way to work. It’s an option, not homework, and it isn’t a prerequisite for anything in this course.

I keep personal folders in mine alongside the business ones — money, health, home — because the things that eat my attention aren’t only the billable ones. That’s a genuine choice, though, and now you know what it means.

What stays out

Anything you’d never want an AI to read. Someone else’s private material, anything you hold in confidence, records that aren’t yours to share. Keep those in a folder you never attach.

If you point at something enormous

Put a short text file at the top of it — fifteen lines, one per folder, saying what’s in each. That’s Anthropic’s own remedy for someone with hundreds of folders, and it takes ten minutes instead of a weekend.

✅ Done when: you can name the one, two or three folders you’re going to attach.

Your AI can only read a file that it can actually get to. There are two ways it can be blocked, and one switch that solves both for most people — so start there and you may be done in thirty seconds.

If your work lives in Google Drive: turn on the Google Drive connector. That’s the whole lesson.

In your AI’s settings, switch on the Google Drive connector. It reaches Google directly and reads your Docs, Sheets, Slides, PDFs, Word files and images straight from the source — nothing has to be downloaded to your computer, and nothing can be out of date, because it reads the live document rather than whatever last synced.

I have never had to mark a single file for offline access. The connector is how I work, and it is the honest recommendation. If your business files are in Google Drive, turn it on and skip to the check at the bottom.

This also quietly fixes the nastiest problem in this whole module, which you would otherwise never spot. A Google Doc is not a file. Drive never downloads one, because Docs, Sheets and Slides don’t exist as files — what lands on your computer is a shortcut of a couple of hundred characters holding an ID. None of your words are in it. It is tiny, so it downloads instantly and looks perfectly healthy. Without the connector your AI opens it and finds an ID number. With the connector, the question never arises.

Dropbox Paper, Box Notes and Notion pages behave the same way, for the same reason.

If your files are somewhere else — Dropbox, iCloud, OneDrive, Box

Those don’t have a connector, so the files really do need to be on your computer. Here is where the second problem bites: all of these services show you file names before they have downloaded the files themselves. Your folder looks completely normal and is mostly empty underneath, and your AI cannot read a name.

⏰ This is the one step that can run for hours, so set it going now and let it work while you watch the rest of the module.

Do it once, to the folder — never document by document:

• Dropbox — right-click the folder → Make available offline. Doing this to a whole folder needs a paid plan; on the free plan you can only mark individual files, so either upgrade or move the folder somewhere local.
• iCloud Drive — right-click → Keep Downloaded if your Mac offers it. Some only offer “Download Now”, and that is fine — you do not need it, because your assistant pulls a file down when it reads one. Only force a download if you later suspect it cannot reach a particular file. What does matter: check System Settings → your name → iCloud → Drive for Optimize Mac Storage — when that’s on, your Mac quietly removes older files again without telling you. iCloud is the only one that does this.
• OneDrive — right-click → Always keep on this device. On a Mac there’s no master switch, so select everything in the folder and do it in one go.
• Box — right-click → Always keep on device, then wait for the green icon.
• pCloud — different in kind. Don’t point your AI at the pCloud drive at all. Use pCloud’s Offline Access to sync into an ordinary folder on your hard drive, and point your AI at that instead.

Google Drive has the same right-click option (Quick Actions → Make available offline, and note that “Download Now” higher up the menu only fetches once). You shouldn’t need it if the connector is on — it’s here for completeness.

“My files are just on my Mac”

Worth fifteen seconds anyway. If Desktop & Documents Folders is switched on in your iCloud settings, then your Desktop and your Documents folder are in iCloud — so the two folders you’re most likely to point at can be full of placeholders while looking entirely normal.

And if it’s on a website rather than in a folder

Then neither of the above applies, and you don’t need to download anything. Your AI can open Chrome and go and look. That’s a later lesson — mentioning it here only so you don’t try to solve a web problem with a folder.

The check — and let your AI do it

Don’t squint at icons. Ask:

It will tell you straight. If you’d rather look yourself: open the folder in Finder, switch to List view, and any cloud icon next to a name means that file isn’t on your computer.

Don’t have your cloud service installed yet?

Everything above assumes the service is on your computer and showing in Finder. If not, install it first — Google Drive for desktop, Dropbox, or OneDrive. iCloud Drive is already built into your Mac; its switches live in System Settings → your name → iCloud, and Apple’s walkthrough is here.

Written August 2026. These steps are the authority — if a button has moved since the video was recorded, trust what’s written here.

✅ Done when: your AI has opened your folder and told you it can read what’s in there — either because the Google Drive connector is on, or because the files are genuinely downloaded.

Updated October 2026. This lesson is text only now, because Claude changed too much for the old recording to help.

Both apps have more than one way in, and one of them is a dead end for people like us. Worth two minutes to get this straight.

In Claude: the desktop app ← this is ours

In October 2026 Claude folded its chat and Cowork into one app. One box now talks with you and works with your real files and folders, does several steps on its own, remembers how you like things, and can drive your browser. Everything in this course happens here, inside the project you’ll make in 1.8.

If your Claude still shows separate Chat and Cowork options, Anthropic hasn’t switched your account over yet. Pick Cowork for now.

In ChatGPT: Work mode ← this is ours

There’s a switcher above where you type, offering Chat, Work and Codex. Pick Work. That switcher catches people out: several of the things this course teaches simply don’t happen in Chat mode, and nothing on screen tells you a switch is the fix.

The website: same account, but it can’t reach your computer

Both also run in a web browser, at claude.ai and chatgpt.com. It’s the same account, so your conversations and projects show up in both. What the website can’t do is work with the files and folders on your computer, and a project tied to a folder on your computer stays on that computer. So for this course, use the desktop app. The website is fine for a quick question from someone else’s laptop.

The developer door

Same engine, most freedom, and it needs developer tools that often refuse to install on an older computer. We are not going there in this course — but not because it is beyond you. I do much of my own work there and I don’t know how to code either; it is simply the rougher door to the same engine. In Claude it’s called Code; in ChatGPT it’s Codex, the third option in that switcher. I’m telling you so that if you land in it by accident, you close it rather than deciding this whole thing is beyond you.

✅ Done when: you’re in the Claude desktop app (on Cowork, if you still see that option), or ChatGPT is set to Work.

Updated October 2026. This lesson is text only now. The clicks below are the ones I walked a member through on her own screen on our October 14 call.

First: make a project from your folder

Claude now works from projects. One project, tied to the folder you chose in 1.5, is what lets every new conversation see your files and your rules.

1. In the Claude desktop app, click Projects, then New project (top right).
2. Give it a plain name, like Business Work, and click Create project. Don’t try to add your folder in this first window. On her screen it wouldn’t let her pick one, and the next step worked fine.
3. In your new project, click Context on the right side (it may say Add context), then Link local folder.
4. Find your folder, click it once, and click Open. Double-clicking opens the folder instead of choosing it.
5. When it asks for access, choose Always allow.

Want to keep work and personal apart? Make two projects, say Business Work and Personal Work, each tied to its own folders.

From now on, start every conversation from inside this project. Click the project’s name and start there. That’s what brings your folder along. A conversation started from Claude’s Home screen doesn’t have your folder, so it has to ask for access, and it can’t see your rules file. When she started one inside her project and typed “good morning,” her morning routine ran straight away, with no permission questions.

Already have conversations you want to keep? Click the three dots next to any conversation in your list and add it to your project. With a conversation open, the small laptop icon at the very top of the screen has an arrow beside it that does the same: Add to project. Once the conversation is in the project, Claude can see your folder there. To find an old conversation, use Search near the top left.

If you connected Google Drive, check it works inside the project too. Type:

In ChatGPT, make a project and use Add folder. If you attach several, one of them is marked primary and new chats start there.

In Claude, a project also keeps its own memory, scoped to that project alone — what it learns about your business stays inside that project. (OpenAI doesn’t document an equivalent promise for ChatGPT projects, so I won’t make one for them.)

Second: the permission setting

It sits in the box you type into. In Claude’s new version there are two:

• Manual (the default) — it stops and asks you before each action. Sensible for your first hour, unworkable after that.
• Auto — this is the one I recommend. It keeps working without stopping to ask, and it still safety-checks each action.

The setting belongs to each conversation, so glance at it when you start a new one. If it’s back on Manual, switch it to Auto.

If your Claude hasn’t updated yet, you’ll see three options instead: Manually approve, Automatically approve and Skip all approvals. Skip is what I used before the change, because the middle one interrupted me every half-minute, and those interruptions are what make people quietly give up. Anthropic’s own words about Skip are worth reading straight: “nothing checks its actions.” The new version doesn’t offer it, so Auto is where everyone lands.

ChatGPT’s equivalent lives in Settings under Permissions, and its widest option is called Full access — the one OpenAI themselves warn “significantly increases the risk of data loss, leaks, or unexpected behavior.” Leave it on the default, where it asks before going outside your folder.

One reassurance that holds no matter which mode you’re in, straight from Anthropic: “Claude always asks before permanently deleting files, in any mode.”

Which one you pick is genuinely yours to decide. If Auto feels like too much on day one, stay on Manual for a day and switch once you trust it. You can change it any time.

One thing worth knowing about Auto: it uses a little more of your monthly plan than Manual, because the checking costs something.

While you’re in there

• In Claude: Settings → General → Run on startup (under Desktop app). Switch it on. Claude can only reach the folders on your computer while the app is open, so you want it already running when you sit down.
• You don’t need to turn on computer control. Mine is off. I’m usually working while it works, and I don’t want my mouse taken over.

Check which model you’re on, and check it again in a month

This one costs nothing and almost everybody gets caught by it. Updating the app does not change your model. A new and better one comes out, the app quietly updates itself overnight, and you carry on using the old one — getting noticeably worse answers, with nothing on screen to tell you why.

The model name sits at the bottom of the box you type into. Click it and pick the newest one offered. Right beside it is how hard it thinks: I use the newest Opus on High, not Medium, and it’s worth a glance each time you start a conversation. I’d put a reminder in your calendar for the first of every month to glance at it — a client of mine had been on the previous model for weeks on a completely up-to-date app, and there was no way for him to know.

Two dialogs, and people mix them up

Your Mac will ask its own permission the first time — that’s macOS, not the app, and if you click “Don’t Allow” the fix lives in System Settings → Privacy & Security → Files and Folders, not anywhere in your AI. Say yes to both.

If it asks for Accessibility and Screen Recording later, you can say no

This is computer control asking a second time, usually weeks in, in the middle of some other task — and because it arrives mid-flow it feels compulsory. It isn’t. Say Deny. What you’re declining is permission to move your mouse and read your screen, and if you say yes it will do that while you’re sitting at the machine trying to work.

The good part: it doesn’t give up, it takes a different route. Told it couldn’t see the screen, mine simply said what the correct answer would look like and asked me to check — and we got the same result a minute later. Turning it on is a real option for people who want it; it just isn’t the price of admission.

Scheduled work, and when your computer has to be on

To set one up, ask from inside your project:

If that doesn’t work, use the scheduling button in the left sidebar.

Since October 2026, Claude’s scheduled jobs run on Anthropic’s computers by default. A job that only needs your connected apps, like Gmail or Google Drive, happens on time even with your laptop closed.

A job that needs the files on your computer or your browser still runs on your Mac, and only while it’s awake with the app open. Neither app can wake a sleeping computer, and closing the lid still puts it to sleep. For those jobs you tell the Mac itself to wake on a schedule, a one-time setup your AI can walk you through. Lesson 2.14 covers it, and it’s written up here: how to make your Mac wake itself on a schedule

The safety net

Saying “undo that” works for changes it made to your files. It isn’t a universal undo for everything that can happen on a computer. Your stronger net is the version history in Google Drive or Dropbox, which is one more argument for keeping your work in a synced folder.

✅ Done when: a new conversation inside your project answers “good morning” without asking for folder access, and the box you type into is set to Auto.

Written August 2026. These steps are the authority — if a button has moved since the video was recorded, trust what’s written here.

A lot of the useful work happens with your assistant operating your browser — your website, your email platform, your course platform, your back office. That’s why this lesson moved from optional to required.

And it has to be Chrome for Claude. Not Safari, not Arc. (Update since this was recorded: ChatGPT’s extension now also lists Edge, Brave, Opera and Vivaldi. If you’re following my demonstration, Chrome is still the one to install.) Keep Safari for your own browsing if you like it — Chrome is the one your assistant works in.

(Update, October 2026: Claude now also has a browser built into the desktop app. If you’ve set up Claude in Chrome, it stays the default, and it’s what I still use, because Chrome is where I’m already signed in to everything.)

Both routes use your Chrome, with you already signed in to your own sites. That’s the whole point of doing it this way.

If you chose Claude

1. Install the official Claude in Chrome extension from the Chrome Web Store.
2. Then, back in the Claude desktop app, turn on the Claude in Chrome connector in Settings → Connectors.

Two things that trip people up here. After installing the extension, quit Chrome and reopen it — it often will not appear until you do. And if you keep more than one Chrome profile, a personal one and a work one, install the extension in both and sign in to your AI in both; it will ask which profile to drive when it matters.

Step two is where people stop, and the extension on its own does nothing without it.

If you chose ChatGPT

Different path, same destination — and it starts inside the desktop app, not in Chrome.

1. In the desktop app, open the Plugins tab and install the Chrome plugin.
2. It walks you through adding OpenAI’s extension to Chrome. Approve the prompts.
3. Switch the composer to Work mode. Browser control doesn’t run in Chat mode at all.

That third step is the one that will otherwise waste your afternoon. In Chat mode you’ll ask it to open a site and simply be told it can’t, with nothing suggesting the mode is why.

Afterwards it can reach for Chrome on its own when a task needs a site you’re signed in to, and you can also point it at Chrome directly by typing @Chrome in your request. Per-site permissions live under Settings → Computer Use, where the four choices are Allow once, Allow for this site, Allow for all sites, and Decline. I use “Allow for all sites”, and I’d suggest you start there too. The narrower settings sound safer, and what they do in practice is stop and ask you every single time it touches a site it hasn’t seen before — often enough that most people quietly give up on the feature. The honest disclaimer: OpenAI flags that setting as elevated risk, and they aren’t wrong. What they mean is a web page carrying instructions written at your AI rather than at you. So the trade is fewer interruptions in exchange for keeping half an eye on what it does on sites you don’t know. If that trade bothers you, Allow for this site is the cautious version, and nothing in this course depends on which you pick.

One thing neither will do

Move your saved passwords, or type a password for you. Handling credentials is against both companies’ safety rules, so asking gets you a polite refusal in your first hour, which is the worst possible time for one. Chrome’s own import moves passwords over in a minute.

✅ Done when: your assistant can open a page in your Chrome for you.

This is the same question you may have run at the start of the session. Run it again now that everything is connected.

Do this:

Open your project and say, out loud or typed...

Three things can happen

1. It names your real folders and describes your business. You’re set up. Done.
2. It can’t see any files. Your files still aren’t fully on your computer — go back and look for cloud icons in Finder.
3. It names your folders but the description is thin, or a bit off. That’s a thin folder, not a broken setup. You’re still set up, and filling it in is what the rest of this is for.

Then keep what it said

Paste the answer somewhere you’ll find it again, and run the same question in three months. The difference between the two answers will tell you more than any progress report. If it looked strange, bring it to the monthly Q&A. And if you and I work together one to one, send it to me.

And the habit to carry forwardAnything you’d otherwise re-explain in every new conversation belongs in the folder. Module 2 is almost entirely about what to put in there.

Michele asked whether your assistant tells you when it has drafted something, so you can edit it before it goes anywhere. It does — and this is the shape of how I actually run a week.

I keep a standing block for social posts. Claude opens the tabs, fills each composer, and stops. I edit what it wrote, click Schedule on one, and tell it to carry that copy across to the rest.

Two habits do most of the work:

• At the end of a working session:

  Create a playbook for yourself based on everything we did today, and schedule anything you can schedule. Prompt me when you need me.

• Every morning: “Good morning, Claude.” It comes back with what is pending — copy waiting on my approval, posts waiting on my yes, things I said I would do and have not.

One more thing that surprises people. I have not used the chat side in six months. Anything a chat can do, Cowork can do — and Cowork can also reach your files and your browser. In October 2026 Claude made that official and merged the two into one box.

Module 2 takes all of this apart properly.