Learn ChatGPT Work in Under 20 Minutes

0:00ChatGPT Work is going to be one of the biggest changes in how you use your computer. It looks the same as regular chat, but it can do things that ChatGPT just simply cannot do. I'm going to show you what those differences are, but let's see an example first. I've attached the invoice folder to this project.

0:20Now I'll give Work a set of rules to follow. From now on, whenever I ask you to process an invoice, rename the PDF using the invoice date and supplier, add its details to the existing expense spreadsheet, and create an Apple Reminders task for any unpaid invoice. Save these instructions in the project's AGENTS.md file

0:40so they apply to future sessions. It's telling us it saved the workflow in AGENTS.md for future invoice processing. In the same project, but this time in a new chat within that project, I've dropped a new invoice called invoice2, and now I can just say: process the new invoice. That's all I have to give it. It will process the invoice, update the spreadsheet,

1:02and now it's opened Apple Reminders and created a reminder for the unpaid invoice. ChatGPT Work is controlling the application via computer use. It's followed the saved rules in AGENTS.md without being told. And in Finder we'll see the renamed invoice PDF. Let's go through the OpenAI product range. We have ChatGPT Chat, which everyone is familiar with,

1:27but then there's Codex, and now there's ChatGPT Work. Chat is for conversations, Codex is for coding, and Work is for long-running tasks. One important difference is that Work uses up your weekly subscription usage allowance, whereas Chat does not. So my advice is to leave your quick questions and brainstorming to ChatGPT Chat in your browser.

1:47Keep your ChatGPT app set to Work.

1:52Next, you need to understand local versus cloud. The choice only exists in the ChatGPT Work desktop app. Local means running and operating on files and folders that are stored on your computer. And any job that is running would stop if your computer was turned off. Cloud is useful because any job that is running keeps on running,

2:13even if you close your computer. But it does have the drawback that it can't see your local files and folders. But if you're using a cloud system like Google Drive, ChatGPT Work in cloud mode can work on those files. Use cloud mode when you want to run a process that doesn't need local file access and needs to run when your computer is turned off. One of the most important features you need to understand is projects,

2:38along with the concept of mounting a folder to that project. You give each project a name and you choose a folder for it to work in, and you can mount multiple folders to a project. This allows ChatGPT to access the files in that folder, work with them, and create new ones. But once you mount a folder, you can create a special instructions file called AGENTS.md,

2:59which is saved in that folder. If you don't choose a folder, ChatGPT makes one for you. On my Mac, it puts it in a ChatGPT folder inside Documents, and there's no setting to change that. So pick your own folder and keep your instructions inside AGENTS.md. If you mount multiple folders, you need to select which folder is the primary folder.

3:19That's where the AGENTS.md file would be for that project: in that primary folder. But this is not a sealed container. It's more of a home base. What really matters is the permission status. When it's set to Ask for approval, it can still read files in other folders without asking. But if it wants to change anything out there, like creating a file,

3:40moving one, or editing one, it asks for your approval first. If you change that setting to Approve for me, and ask it to operate in folders outside that project, it'll be able to do that without asking you first. Mounting a folder is not the same as attaching a file. Everyone has used that plus icon to attach an image or a file to a chat thread.

4:01You're essentially telling the AI: use this for this conversation. Another chat in the same project might find that file through the earlier chat. You can't count on it. If you want every chat to have it, ask Work to save it in the project's folder. One of the reasons why using folders is so useful is because they sit locally on your machine, and if you decide to move to the Claude or Gemini AI platforms,

4:24your work is not stuck inside ChatGPT. Your work is in those folders. In the first example, we asked ChatGPT Work to create an AGENTS.md file from our instructions. Here's that file. It's just a normal markdown file of instructions for the agent. In this case, it tells Work how to rename invoices, update a spreadsheet, and create reminders.

4:48You can edit the instructions yourself, or you can ask Work to update them for you. So the thing to understand is there's actually two levels. There's a global agents file, which provides instructions for all of your projects, and then there's a project agents file, which contains instructions specific only to that project. You can even put an agents markdown file inside a subfolder,

5:10so different parts of a project can have their own instructions. To find the global agents markdown file, open Settings, go to Personalization, and click Codex Instructions. This is where you set your rules for how you want Work to behave across all of your projects. And a rule can be something really simple. Say, how you want files named: I want images named

5:32so I can tell what's in them from the file name alone. And Work actually looks at the picture to decide what to call it. Let's create an agents markdown file for our marketing project. Unlike the global instructions, these rules apply specifically to this project. We don't need to write the file manually; we can simply ask Work to create it for us.

5:52The one thing I would say, though, is don't bloat the file. You don't want the file becoming enormous. Every instruction takes up context, and there's a size limit too: anything past it simply isn't loaded. The better approach is to use the AGENTS.md file as a routing map. Instead of putting all your detailed procedures in one file,

6:13split them into separate markdown files and tell the AI when to read them. Here we're asking Work to create three supporting documents. One provides background on the project, another defines the marketing research workflow, and the third establishes standards for our deliverables. Then we update AGENTS.md so its routing section says when to read each one.

6:35This is often called progressive disclosure. We'll start a new conversation in our marketing project and ask Work to research three effective ways to promote a business on TikTok. Notice that we haven't mentioned the agents markdown file or any supporting document in our prompt. We can also verify whether the instructions have been followed, rather than simply assuming that they were.

6:59Now, let's talk about plugins. Plugins connect ChatGPT Work to the apps you already use, like Gmail, Google Drive and Slack, so it can find things in them and get things done in them for you. If the plugin you want is available, connecting it is a really simple operation. You find it, you connect it, you sign in, and it's done.

7:19A plugin can include apps, skills, or both. So when you turn one on, it can also bring skills with it, which tell ChatGPT how to use that app properly. And some plugins are mostly skills. OpenAI's data plugin, for example, is about twenty skills for things like building reports and dashboards. So how exactly does ChatGPT decide to use a plugin? ChatGPT uses your installed plugins automatically,

7:44when it thinks it's relevant. So if you've connected Gmail and you say: I think I've got a flight booked next Wednesday, is that right? it might go and search your email for that flight booking, without you ever mentioning email. Or you can be explicit. You type the @ symbol and pick the plugin, so @Gmail.

8:01You can just name it in your prompt as well. You can build it into your AGENTS.md, so for certain kinds of tasks it always uses certain apps, or you can put it inside a skill. And what if the tool you use doesn't have a plugin? Are you stuck? No, absolutely not. A lot of tools now have what's called an MCP server. So you search for the name of your tool plus MCP,

8:24go to its documentation, and find the MCP server URL. Then, in ChatGPT on the web, go to Plugins, press the plus button, choose Add custom MCP server, paste in that URL, sign in, and you're connected. A couple of things to know: it's web-only at the moment. If you're on a Business plan, only your admin can do it, and on Pro it's limited to reading

8:45and fetching information rather than making changes. Skills are one of the most useful features inside ChatGPT Work, and they're a lot simpler than they sound. A skill is a set of instructions for doing a particular job the way you want it done. And if we click on one, you'll see it's essentially just a text file:

9:03a markdown file called SKILL.md. There are a few lines at the top, a name and a description, and then the instructions underneath. The description is what tells ChatGPT when to use it. When ChatGPT uses a skill, it's just reading that text into the conversation. There's really no difference between that and you typing all those instructions yourself and hitting go.

9:25It just saves you doing it every time. So how do you make one? The easiest way is just do the job with AI first. Do a task, go back and forth until it's exactly how you want it, and then, at the very end, just say: make me a skill on what we've just done.

9:42ChatGPT has a built-in skill creator, and it will write that skill for you. Your skills then live in the sidebar, under Plugins, in the Skills tab, and that leads to something really important. Skills are why you shouldn't finish things off yourself. Stay in the chat and tell the AI to make those last little tweaks,

10:01even the small ones, until you're completely happy. Then, when you make the skill, all of those little tweaks are in it, rather than missing because you did it yourself. And skills are not one and done. You might use the skill again the following week, and it doesn't quite do what you wanted. That's fine. Get the AI to fix the result, but then tell it:

10:18look, we've had to make these tweaks. Can you update the skill to account for this? Skills can be edited, amended, extended, or completely changed over time. Let's talk about scheduled tasks, because in my view, this is what really unlocks the power of ChatGPT Work. Everything we've talked about so far is you sitting there,

10:38asking it to do something. Scheduled tasks are where it starts doing the work without you. There are two kinds, and it's worth getting the difference clear. A schedule is: run this every day at nine, or every Monday morning. It runs whether or not anything has happened. A trigger is: run this when something happens. So when an email arrives from a particular person,

10:58or with a particular subject, it kicks off straight away. Triggers include things like a new email in Gmail or a message in a Slack channel, and some other plugins can start tasks too. If your scheduled task needs files on your computer, like my audio notes, your computer needs to be on, and the desktop app needs to be running when it goes off. If the task only needs your connected apps, like Gmail or Drive, you can run it in the cloud,

11:23and it doesn't matter whether your laptop is open or not. A couple of practical things: you'll find all your tasks on the Scheduled page. You set up triggers on the web or on your phone, not in the desktop app, which can show them but can't create them. And there's a limit on how many tasks you can have running at once,

11:41depending on your plan. Let's talk about browser use and computer use. In the desktop app, Work has its own built-in browser that opens up beside your chat. It's a fully functional browser. You've got tabs, history, you can sign in to sites, it's got autofill, password management, even extensions. It's also where your files preview. When Work makes a document or a presentation,

12:05you see it right there next to the chat. The big unlock here is you don't have to do the browsing yourself. You type @browser and you tell Work to do it. One example I really like is keeping an eye on a page. Say you want to keep up with what a competitor is releasing.

12:20You tell it: use your built-in browser to open this release notes page. Read the page directly, don't rely on search results, and keep a CSV file in this project with every new release. And it goes off, reads the page, and builds that file for you. And there's also this really handy feature where you can point at things.

12:38In the browser, there's an Annotate button. You click on something on the page, or drag over an area, leave a comment on it, and then ask ChatGPT to do something with it. And you can do the same on a file it's made. So you can point at one chart on one slide and say: change this.

12:54You can also sign in to your apps in its built-in browser. Your saved details fill in, but you press login yourself. You can connect Work to your own Chrome with the ChatGPT extension and type @Chrome, and it works in the apps you're already logged into in your browser. And that brings us to computer use.

13:12So there's really three ways that Work can act on things. A built-in browser, with @browser. Your own Chrome, through the extension, with @Chrome. And computer use, with @computer or the name of the app, which is for normal desktop apps, things like Excel, PowerPoint or Photoshop. With computer use, it's literally looking at your screen,

13:31clicking and typing like a person would. On a Mac, you'll need to give it screen recording and accessibility permissions, and on Windows, it takes over your mouse and keyboard while it works, so bear that in mind. Turn computer use on from Plugins, and you install the Chrome extension from Settings, under Computer use. Next up, Sites.

13:53ChatGPT Work can build a website and host it for you on OpenAI's servers. It's a really good way of sharing information, much better than sending someone a PDF. You make a site from a chat in Work. You describe what you want and mention that it's a website, or type @sites. The Sites button on the left-hand side isn't where you create them.

14:11It's where you find and edit the sites you've already made. So, say you run a shop. You could then say: build me a website for my shop with a home page, our five products, delivery and returns, and a contact form. You don't need to know anything about code or hosting. It builds it, shows you a preview, and you just keep telling it what to change:

14:29make the header smaller, add a gift guide page, whatever you need to change. When you're happy, you publish it. If it's public, people visiting don't need a ChatGPT account. It just looks like a normal website. By default, you get an address on chatgpt.site. You can use your own domain. ChatGPT won't buy the domain for you, so you still buy that yourself.

14:51For a business, though, I think one of the best uses is internal dashboards. Information for your people in your company. Something someone can update, or that gets updated regularly, and you control who sees it. By default, it's just you. You can share it with selected people by email, with anyone in your workspace if you're on a Business or Enterprise plan, or with anyone on the internet.

15:13Just be careful with that last one. And the people you share it with can either view it or edit it. Sites is on the paid plans, not Free or Go, and depending on where you are, publishing publicly might not be available yet. When it launched, it wasn't available in the UK or Europe, so check your account. Finally, a few things that make working

15:36with ChatGPT Work day to day a lot smoother. It often makes a plan, and you'll see that plan in the chat, along with the sources it's using and the files it's making. If you look at the plan and don't like it, you can just chime in and tell it to change something. Now, you don't want to read every plan.

15:52I give it a job and I go and do something else. Usually, I've got two to four tasks going at once. As one's working, I'll start another, or check what a finished one has done. It's like having a co-worker who gets on with things while you do something else. And while it's working, you don't have to just sit there in the dark.

16:10You can click to expand and see what it's doing, every step it's taking. Then there's steering. Everyone knows how to prompt, but a lot of people don't realise you can send another message while it's still working. When you do, you can choose to steer, which adds your message to what it's doing right now,

16:26or queue it, so it picks it up once it's finished. There's also fork. If you type slash fork, it copies the whole conversation into a new chat, with all the context, so you can go off in a different direction without messing up the original. If you archive a chat, it's not deleted. It just disappears from your sidebar. You'll find it in Settings, under Archived chats,

16:47where you can bring it back or delete it properly, and it still shows up in search. If you leave tasks running on your Mac while you step away, turn on Prevent sleep while running, in Settings, under General. Just know it won't stop a MacBook sleeping when you close the lid, unless it's plugged in with an external screen. And one setting to check if you're using this for business:

17:08Improve the model for everyone, under Data controls. On Free, Plus and Pro, it's on by default, so your chats can be used to train the AI model. I'll turn that off. And one last tip, which I learnt the hard way while making this video. If you ask ChatGPT about its own features, it doesn't always get them right.

17:26A few times it told me one thing, I pushed back, and it changed its answer. So if something it tells you about how the app works sounds a bit off, question it, and check. So keep an eye on the plan when it matters, steer it when you need to, fork it when you want it to branch off,

17:40and don't take everything it tells you as the complete gospel.