How to Use ChatGPT Work: Access, Setup and Your First Task

ChatGPT Work is the mode where ChatGPT stops answering and starts doing: It takes a whole task, works on your real files and tools, and hands back a finished deliverable. Knowing that it exists is one thing. Actually setting it up so it works is another, and that's where most people stall. This is the practical walk-through, from getting access to briefing your first real job.

If you want the full picture of what it is first, our guide to ChatGPT Work covers that. This piece is the how-to: Access, setup, and the first task.

Step 1: Get the desktop app

The first thing to know is that ChatGPT Work is not really a web feature. Work is available to every plan on the desktop app, and to Plus, Pro, Business, Enterprise and Edu on web and mobile, but the web version doesn't carry the full feature set and will keep nudging you to open the desktop app anyway.

So the setup starts with downloading the desktop app for your Mac or Windows machine and signing in. Even a free account gets Work this way. The reason the desktop app matters so much becomes clear in the next step: It's the only place you get both of Work's modes, and one of them is the whole point.

Step 2: Understand local mode versus cloud mode

Work runs in two modes, and choosing the right one is the difference between it being useful and being frustrating.

Local mode runs on your own computer. It reads the files on your disk and gives you the browser built into the desktop app. This is the mode that lets Work do the thing that regular chat can't: Work directly with your actual files, no uploading, no size ceiling.

Cloud mode runs on OpenAI's servers. Its one real advantage is that a job keeps running with your machine switched off. The trade-off is that it can't touch your local files, so it's for jobs that live entirely in connected apps and the web.

For most knowledge work, local mode is the one you want, because the point is usually to work on your own documents. The one catch is physical: A local job needs your computer awake while it runs, since the files live there. Our breakdown of ChatGPT Work versus chat goes deeper on what each mode unlocks.

Step 3: Set up a project and connect your folders

Here's the habit that separates people who get value from Work from people who don't: Every task should live in a project. A project is the container that tells Work which folders on your machine it can read from and write into.

To set one up, start a task, create a project, name it, and select the source folders it should have access to. A project can hold more than one folder, but only one primary folder, which is where it looks for standing instructions. Once your folders are connected, the right-hand panel gains a Files tab showing everything in them, which matters because Work is usually creating files as well as reading them, and you want to see what it produces. This folder-scoped access is deliberately similar to how Claude Cowork stays scoped to the folders you connect.

Step 4: Connect the tools you use

Work reaches your other software through installable packages that ChatGPT calls plugins, and each one can hold two different things. Apps connect Work to a tool, the way connectors wire it into the software you already run. Skills teach it a repeatable workflow. An app tells it where to work; a skill teaches it how.

So go through the plugins for the tools you actually use, your email, your drive, your CRM, and connect them. Where a tool has no plugin at all, there's a shared standard called MCP that lets it connect anyway, though that's a slightly more advanced setup. Before you give anything write access, open it and read its list of actions, because that's where the surprises live.

Step 5: Brief your first whole job

Now the part that changes how you work. The skill isn't writing a clever prompt, it's describing a whole deliverable. The test: Can you say what the finished thing is? "What should go in a competitor analysis" is a chat question. "Build a competitor analysis from these five files, one tab per competitor, with a summary tab at the front" is a job for Work.

So give it a real task with a clear output. Point it at the folder, describe the deliverable, and add any check you want built in, like "flag anything you're not sure about rather than guessing." When Work shows you its plan, read it. Stopping a job at minute two because the approach is wrong costs you far less than discovering it twenty minutes in. Then let it run. A real task can take minutes rather than seconds, because it's doing the work rather than describing it.

Step 6: Stop repeating yourself with standing instructions

Once you've run a few jobs, you'll notice you keep telling Work the same things: Where a file lives, how you like output formatted, a rule it should always follow. That's the signal to write it down once. A plain instructions file in your project's primary folder holds persistent context that Work reads at the start of every session in that project, so it stops needing the reminder.

The rule of thumb is simple: If you find yourself typing the same instruction twice, it belongs in that file. It's the same idea as ChatGPT Projects for everyday chat, or a standing brief for a new hire, the context you'd otherwise re-explain, captured once.

What to try first

Pick something real but low-stakes for the first run, so you learn the tool on a job where a mistake costs nothing. Good first tasks: Turn a folder of receipts or invoices into a tidy spreadsheet. Pull several documents into one written summary. Reformat or rename a batch of files by a rule. Draft a report from data you point it at.

A concrete first run makes the pattern click. Point a project at a folder of receipt images and ask it to read each one, build a spreadsheet with a row per expense, and sort the images into subfolders by month. It'll make its own plan, work through the files, and hand you the finished spreadsheet with the images filed, flagging anything it couldn't read cleanly rather than guessing. That single job shows you the whole loop: A real deliverable, produced from your own files, with a check built in.

And when you're ready for it to run without you, a scheduled task can fire one of these jobs on a cadence, though note that scheduled tasks sit on the paid plans. Start with one job, prove it, then hand over the next. One more thing to expect: File creation draws on your usage allowance harder than a normal chat does, so a heavy run costs more than a quick question, which is another reason to start small.

Getting a team past the setup and into actually handing over whole jobs, briefing them properly and checking what comes back, is the part that needs teaching, and it's exactly what our Cowork training programme does, function by function, on the tools you already run. If you want to be the one on your team who makes this work rather than the one who gave up at setup, that's where to start.

Common questions

How do I access ChatGPT Work? Download the ChatGPT desktop app, sign in, and use the Work mode. Work is on every plan through the desktop app, which is where you get the full feature set; the web version is more limited and will prompt you to open the app anyway.

Do I need a paid plan to use ChatGPT Work? No. Work is available on the free plan through the desktop app, with usage limits. Paid plans add web and mobile access, higher limits, and features like scheduled tasks.

What's the difference between local and cloud mode? Local mode runs on your computer and can read your own files, but needs the machine awake. Cloud mode runs on OpenAI's servers and keeps going with your machine off, but can't reach your local files. Most file-based work wants local mode.

How do I give ChatGPT Work my files? Set up a project and connect the folders it should read from and write into. It only works inside the folders you connect, and a Files tab shows you everything in them, including what it creates.

Why is my ChatGPT Work task taking so long? Because it's doing the work rather than answering. Runs of several minutes are normal for a real job. Watch the plan when it appears, and stop it early if the approach is wrong.