A scheduled task is a prompt that runs itself, on a cadence you set
You write out what you want done, once, and say when it should happen. Every weekday at eight, every Friday at four, the first of the month. After that it goes on its own and leaves you the result.
The difference from a calendar reminder is the whole thing. A reminder tells you to write the weekly summary. A scheduled task writes it, and it’s already done next time you go looking.
Claude and ChatGPT both keep them in the same place: Scheduled, in the left-hand sidebar. You can also just say it in a chat: “set this up as a scheduled task to run every Monday at nine.” It’ll build it from that. Then open Scheduled and check the task it made before relying on it.
Build it as a skill first, then make it a scheduled task
Putting a time on something doesn’t improve it. Whatever it does while you’re sitting there watching is exactly what it’ll do at seven in the morning when you’re not, so you want to have seen it work first.
That’s what building it as a skill gets you. Run it by hand a few times, fix what’s wrong, and only once you’re happy with what comes back do you put it on a timer.
You keep the hand-run version either way. When something comes up out of cycle, a client asking on a Wednesday for the thing you send on Fridays, you run the skill and you have it.
Claude and ChatGPT handle unattended actions differently
The whole point of a scheduled task is that you’re not there when it runs. So when the job gets to something that matters, sending an email, changing a record, posting a message, something has to happen and nobody is around to be asked. Claude has one answer to that. ChatGPT’s depends on where the task runs.
Claude runs it. A new task is set to Automatically approve, and the dialog tells you what that means: it will work and use your connectors without pausing. There are two other settings, and the wording is theirs:
Manually approveClaude pauses so you can approve each action.Automatically approveClaude keeps working without asking about every step, while checking actions for safety. This is the default.Skip all approvalsClaude never pauses, even for unsafe actions.
ChatGPT runs it on your computer, and may stop in the cloud. A task that runs in a project on your computer doesn’t stop to ask. What it’s allowed to do is decided before it starts, by the permission setting you already use in ChatGPT: look but change nothing, change files inside the project folder only, or full access. On a work account, your admin can switch that off, and then the task asks the way a normal chat would. A task that runs in the cloud is more careful with your connected apps. Anything that sends a message or changes something outside ChatGPT may need approval, and “if approval is required, the task pauses until you review it.”
The practical difference is what happens when an unattended run reaches an action that needs oversight. A Claude task on Automatically approve carries on, and so does a ChatGPT task running on your computer. A ChatGPT task in the cloud may stop and wait for you.
It isn’t really a question about tasks. It’s whether you’re happy for it to use your connectors while you’re not watching. If you are, automatic is fine. If you’re not sure, press Run now and watch the whole thing happen, including what it touches and where the output ends up. Until you’ve done that, keep anything that leaves the building at a draft and press send yourself.
A good first one to build is the same one I set up on training days. It reads your inbox and your calendar and tells you what’s sitting unanswered, what you’re still waiting on from other people, and what’s coming up that you should have prepared for. It writes the answer into the chat and sends nothing anywhere, so there’s nothing to be nervous about while you decide whether you trust it.
Set it to run in the cloud, or on your laptop
Every task runs in one of those two places, and the choice decides what it can reach.
- In the cloudRuns at the time you set whether your laptop is on, off or in a bag. It reaches your connected apps and anything saved to your account, and it can't see a folder on your machine.
- On your laptopDoes everything the cloud one does, and reaches your second brain and your other local files as well. The catch is that it only runs while the machine is awake and the app is open.
The rule that settles it: if the task needs to read or write your second brain, it has to be a local one. If it doesn’t, it will work fine in the cloud.
Which means if you’ve got a second brain sitting in your Documents, start local. Most of what you’ll want a task to do is going to want that folder sooner or later.
If you’re going local, turn on the setting that keeps your computer awake, and leave the lid up. Closing the lid puts the machine to sleep, and a sleeping machine runs nothing.
With Claude, a local task that misses its slot starts when you next open your laptop, so a 9am task can run at 11pm. ChatGPT does both as well. In its Scheduled list, a task with a cloud icon runs in the cloud; one without it runs in a project on your computer, and needs the machine on and the ChatGPT desktop app running.
Send the output somewhere you’ll actually see it
Each run leaves a result you can review from the task’s history or associated conversation. That works for about a week. Then you stop opening the list, and the task carries on running into an empty room.
So decide where the result should go while you’re building it, not after: a draft in your inbox, a file in the folder you work out of, a message in a channel you’d have read anyway.
Don’t treat a completed run as proof that the useful job got done correctly. If the task matters, open a run every so often and read what it actually produced or changed.