Most people using ChatGPT every day have never opened the Scheduled panel. It sits in the sidebar, it is now available on the free plan, and it turns ChatGPT from something you visit into something that shows up. Not an agent framework, not a workflow builder, just a prompt that runs on a clock.
The gap is worth closing this week. A scheduled task takes about ninety seconds to create and removes a recurring twenty-minute job from your calendar permanently. Below is what the feature actually does, what each plan allows, the prompt structure that produces useful output instead of vague summaries, and where it quietly fails.
What is a ChatGPT scheduled task?
A scheduled task is a saved prompt that ChatGPT runs automatically at a time you choose, then delivers the result as a notification and a new chat. You write the instruction once, set the cadence, and ChatGPT executes it without you opening the app. It can use the same tools an interactive chat can, including search and connectors.
That last part is what changed the feature from a novelty into something usable. An early scheduled task could only reason from the model's own knowledge, which meant it produced generic text on a timer. A task that can search the web, read your connected Gmail or Slack, and then write is doing real work.
You create one by opening Scheduled in the sidebar, or by writing a prompt that contains an explicit schedule and confirming when ChatGPT offers to save it. Both routes land in the same panel, where you can edit, pause, or delete.
How many scheduled tasks does each ChatGPT plan allow?
Active task limits are set per plan, and they are low enough that you have to choose deliberately. Free and Go accounts get three. Plus gets five. Business and Edu get ten. Pro and Enterprise get fifteen. Free accounts are also restricted to flexible scheduling windows rather than exact times.
Here are the constraints worth writing down before you plan anything:
Active scheduled task limits by plan
- Free and Go: 3 active tasks, recurring no more than once per day
- Plus: 5 active tasks
- Business and Edu: 10 active tasks
- Pro and Enterprise: 15 active tasks
- Paid plans: recurring tasks as often as once per hour, with exact delivery times
- Free and Go: flexible windows only, such as morning, afternoon or night
- Event-triggered tasks: require ChatGPT Work and an eligible paid account, not available on Free or Go
The practical consequence: on Plus you have five slots, so a task has to earn its place. A daily digest you actually read earns a slot. A weekly "inspire me" prompt does not.
Why do most scheduled tasks produce useless output?
Because people write them like a chat message instead of a standing instruction. In a live chat you can correct a vague answer in one follow-up. A scheduled task has no follow-up. Whatever the first pass produces is what lands in your notifications, so every constraint you would normally add through conversation has to be in the original prompt.
The failure looks the same every time. Someone writes "every morning, summarise the AI news" and receives a shapeless paragraph that could have been written on any day of any month. Nothing in that instruction tells the model what counts as news, what to ignore, how long the output should be, or what the reader is going to do with it.
A scheduled task prompt needs five things a chat prompt can leave implicit: a defined source scope, an explicit recency window, a filter for what to exclude, a fixed output format, and a rule for what to do when there is nothing worth reporting. That last one matters more than it sounds. Without it, the model will manufacture content on a slow day rather than tell you the day was slow.
The scheduled task prompt template that works
Copy this, replace the bracketed parts, and paste it into ChatGPT as a new message. It contains the schedule inline, so ChatGPT will offer to save it as a task.
Try this prompt
Every weekday at 8:00am Hong Kong time, do the following.
SCOPE: Search for news published in the last 24 hours about [your topic, e.g. AI tools for marketing teams]. Prioritise primary sources: official company blogs, product changelogs, and filings. Deprioritise aggregator sites and opinion posts.
EXCLUDE: funding rounds under [US$50m], speculation about unreleased products, and anything I could not act on this week.
OUTPUT: Maximum 5 items. For each item give me exactly three lines:
1. A one-sentence factual summary, no adjectives.
2. Why it matters to [your role, e.g. a content marketer running a 4-person team].
3. The source link.
Then add one closing line: "What I would do today: [one concrete action, or 'nothing']".
IF NOTHING QUALIFIES: Reply with exactly "No qualifying items in the last 24 hours" and stop. Do not pad the output.
TONE: Direct. No preamble, no "here is your daily briefing" opener.
The instruction that changes the output most is the last conditional. Give the model explicit permission to return nothing and it starts filtering honestly. Leave it out and you get five items every day whether five items exist or not.
How do you share a scheduled task with someone else?
Sharing sends a snapshot of the task rather than access to your account. Open Scheduled, find the task, open the more options menu, select Share, and copy the link. The recipient sees the task title, instructions, schedule, and original time zone, and can create their own separate copy that runs on their account with their connectors.
This is the underrated half of the feature. A prompt you spent forty minutes tuning is an asset, and until now the only way to pass it to a colleague was to paste raw text into a message and hope they rebuilt it correctly. A shared task link carries the schedule and the time zone with it, so the copy behaves the way yours does.
One caveat before you share: the recipient's copy runs on their plan, with their connectors and their limits. A task that reads your Gmail will not read theirs unless they connect their own. Check that the task still makes sense when the data source changes hands.
Three scheduled tasks worth one of your five slots
Slot economics force prioritisation, so here are three that consistently justify themselves for people doing content, operations, or client work.
The competitor changelog watch. Weekly, on Monday morning: search the product blogs and changelogs of three named competitors for anything published in the last seven days, and report only shipped features, not announcements of future plans. This replaces a browser-tab ritual most people abandon after two weeks.
The inbox-to-decision digest. Daily, connected to your mail: list every message from the last 24 hours that contains a question directed at you and is still unanswered, with a one-line summary each. This is not an inbox summary. It is a list of things blocked on you, which is a much shorter and much more useful list.
The Friday self-audit. Weekly, Friday afternoon: given a list of your current projects, ask three specific questions about what has not moved this week and what the likely reason is. The value is not the model's insight. It is the forced fifteen minutes of review that otherwise never happens.
Notice what these have in common. Each one produces a short list of specific items, each one has a clear next action, and each one replaces a habit you already tried and failed to maintain manually.
Where scheduled tasks break down
They are genuinely useful and genuinely limited, and knowing the limits saves you from building on the wrong foundation.
The output is a chat, not a database. Each run creates a new conversation, so there is no accumulating record you can query later. If you need trend data across weeks, a scheduled task is the wrong tool and you should be writing to a spreadsheet through an automation platform instead. We covered where that threshold sits in our comparison of what no-code automation platforms actually cost.
Quality drifts as the source landscape changes. A prompt tuned in September against a particular set of sources will slowly degrade as those sources change format or go quiet. Scheduled tasks are not fire-and-forget. Review each one monthly and either retune it or kill it.
Long-running instructions inherit every weakness of long context. A task that asks for analysis across a large body of pasted material will produce weaker output than the same request in a fresh, focused chat, for the same reason described in our piece on why AI answers get worse the longer a conversation runs. Keep tasks narrow.
And the limits are real. Five slots on Plus is not an automation platform. It is five good habits, automated. Treated as that, it delivers. Treated as a replacement for a workflow tool, it disappoints.
Set one up in the next ten minutes
Pick the recurring information job you most often skip. Not the most important one, the one you skip. Write it using the template above, paste it into ChatGPT, confirm the save, and then leave it alone for a week before judging it. The first two runs are usually mediocre, because you learn what the prompt was missing only by reading what it produced.
Then tune once, and share the link with the one colleague who has the same job. Good automation spreads sideways faster than it scales upward.
Tools do not become useful at the moment they launch. They become useful at the moment someone sits with you and works out which of your actual jobs they replace. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Now that you have the technique, the next step is building it into a workflow that runs reliably every time. We'll walk you through every step, from tool setup to workflow design and deployment.
