Schedule Recurring Tasks in ChatGPT and Claude: A Migration Checklist

If you paste the same request into a chat every morning or every Monday, you can schedule recurring tasks in ChatGPT and Claude so the request runs without you. Setting the timer is the easy part. The vendors' own documentation puts the risk somewhere else: in the move from a chat you watch to a run nobody watches. This guide walks one routine through that move, from choosing where it runs to checking that it actually did the work.

Both products document the feature. OpenAI's ChatGPT automations guide describes tasks that run in the background, and Anthropic's Help Center article on recurring tasks covers scheduled tasks in Claude Cowork and in a newer Claude experience that combines chat and Cowork. Anthropic says that experience is rolling out to Pro and Max plans, and on Team and Enterprise plans it is available where unified Chat and Cowork is turned on. Depending on your account, the feature may sit under Cowork or simply under Claude. Menus differ by rollout, so I describe the decisions rather than click paths.

First, is this request worth scheduling?

OpenAI's Get started with ChatGPT Work guide says to use scheduled tasks when you want ChatGPT Work "to repeat, monitor, or refresh something over time." Anthropic's examples sit in the same territory, starting with daily briefings that summarize Slack messages, emails or calendar events from the past 24 hours.

Google adds a useful boundary in its Gemini scheduled actions help page: rapidly changing data such as stock prices won't be the latest by the time you receive the result. My rule of thumb from that: schedule summaries, wrap-ups and "what changed since last time" checks. Keep anything that depends on the current minute in a live chat.

Step 1: Decide where the task will run

This decision determines whether your 8:00 briefing arrives when your laptop is closed. The question is simple: does the task need files or apps on your computer?

Anthropic's page says Claude scheduled tasks run remotely, so they keep their cadence while your computer is asleep or the Claude Desktop app is closed. If you read an older guide that says otherwise, it is describing earlier behavior: XDA reported in March 2026 that scheduled tasks ran only with the computer on and the Desktop app running. The current official page is the one to follow, with two exceptions. A task that requires local files or apps "will only run locally." And on Pro and Max plans, Anthropic says scheduled tasks created before October 6 that run on your computer keep running there. If you set up a Claude task before that date, check where it runs instead of assuming it moved to the cloud.

The two Claude Code options are for people who already use Claude Code. If that is not you, the regular Claude app or ChatGPT is the simpler path.

What happens when a local run is missed

Local runs have one more failure mode to plan for. Anthropic's Claude Code Desktop scheduled tasks doc says that when the computer wakes, runs missed in the last seven days collapse into a single catch-up run for the most recent missed time. The same page warns that "A task scheduled for 9am might run at 11pm if your computer was asleep all day," and advises putting time guardrails in the prompt. If your morning summary only makes sense in the morning, a late catch-up can deliver a stale or misleading result. That is a reason to prefer a cloud option whenever the task does not need your files.

Step 2: Rewrite the prompt so it stands alone

The request you paste every day leans on context: the earlier conversation, the file you attached last week, your memory of what "the usual" means. An unattended run should not depend on any of that. OpenAI's guide says durable instructions belong in the task prompt or an attached skill, and the source material the task needs should sit in a project, upload or connected service it can reach. Anthropic's cloud scheduled tasks doc makes the same point more bluntly: "the prompt must be self-contained and explicit about what to do and what success looks like."

A standalone prompt answers five questions:

1. Sources: which connected service, project or upload to read, by name.
2. Window: what time period counts as new.
3. Output: the format and length you want back.
4. Success: what a good run contains, and what to say when there is nothing new.
5. Guardrails: what not to do (send, edit, delete) and what to do if the run fires at an unexpected time.

OpenAI's own example scheduled prompt shows the shape: "Every Monday morning, review new updates from @Slack and @Google Drive for [project]."

Here is an illustrative example of the rewrite. I did not schedule or run it, and it has no recorded output:

The last sentence is the time guardrail. On a local task, it tells you when a catch-up run fired hours after it was due.

Step 3: Test it by hand, then do a dry run

OpenAI is direct about the order: "Before you schedule a task, test its prompt in a regular web chat." Paste the rewritten prompt into a fresh chat with no earlier context. If the answer is wrong or vague there, expect the scheduled run to be no better, with nobody watching.

Local Claude Code tasks add a second check. Anthropic's Desktop doc says that a task in Manual permission mode stalls when it reaches a tool it has no permission for, and waits until you approve it. Its fix is to run the task once now, watch for permission prompts and choose "always allow" for each one. Do that dry run while you are at the computer, and read each prompt before you approve it: every "always allow" is a permission the unattended runs will keep.

Step 4: Narrow permissions and connectors

In chat, you see each step. In a scheduled run, you don't. OpenAI's guide says scheduled tasks run unattended with your default sandbox settings, so whatever your defaults allow, the task is allowed too.

Anthropic's cloud scheduled tasks doc is more specific. Every connector on your account is included in a new task by default, and Claude can use an included connector during a run, writes included. You don't pick a permission mode for these tasks: a run on Anthropic's cloud starts in auto mode, where a classifier model allows or blocks actions, unless auto mode is unavailable. A run still stops and waits for you when an action matches an ask rule or when blocks pile up past auto mode's thresholds. Anything the task does through your connected GitHub identity or connectors "appears as you."

A summary task that only needs to read Slack should not carry a connector setup that lets it post to Slack or edit Drive files. I recommend removing every connector the task does not read from. Keep my example prompt's instruction against posting, editing or sending anything too, but treat it as an instruction the model is asked to follow, not a permission boundary: what a run can actually do is set by the connectors, linked accounts and sandbox settings you leave in place. For a broader framework on scoping what an agent can touch, see the agent permissions playbook.

Step 5: Set the schedule, including a Sunday-to-Thursday week

In ChatGPT, OpenAI distinguishes a standalone scheduled task, where each run starts from the saved prompt, from a scheduled task in a chat, which returns to the same chat with its existing context. Since Step 2 made the prompt stand alone, I recommend the standalone kind.

Claude's presets, per Anthropic's Help Center, are hourly, daily, weekly, on weekdays, or manually.

A note for anyone whose work week runs Sunday to Thursday, as it does in Israel. The Claude Code Desktop doc defines its Weekdays preset as "same as Daily but skips Saturday and Sunday." That skips your Sunday and runs on your Friday. My read is that a Sunday-to-Thursday week needs a custom schedule rather than that preset. ChatGPT offers custom schedule controls for custom cadences, and its automations page also mentions advanced recurrence rules in the RRULE format (RFC 5545). In that standard, a Sunday-to-Thursday rule would look like FREQ=WEEKLY;BYDAY=SU,MO,TU,WE,TH; treat this as an illustration of the syntax, not a tested task. The pages I relied on do not define which days the Claude app's "on weekdays" preset or ChatGPT's presets include, so check the day list in the schedule screen before you trust a preset.

Step 6: Check results, not the status light

My recommendation for any scheduler: treat a successful status as proof that the run happened, not that it did the job. Anthropic's cloud scheduled tasks doc draws the same line for its run history. A run marked Succeeded only means the session ended without Claude Code reporting a failure; blocked network requests, missing connector tools and work Claude couldn't finish show up in the run's transcript, not on the row.

A run can finish cleanly and still read an empty channel, miss a disconnected source or summarize the wrong week. In ChatGPT, runs with findings appear in the Scheduled view, and OpenAI says an unread indicator shows when a run needs your attention. OpenAI's guide also advises reviewing the first few runs and adjusting the prompt, tools or cadence if the results are too broad.

For the first two runs, open the result and compare it against the sources yourself. Does each bullet link somewhere real? Did "nothing new" mean nothing happened, or that the connector lost access? Zapier's author, writing about ChatGPT's event-triggered tasks specifically, called that feature "still a little glitchy" and recommended "trialing a few tasks and monitoring closely" (Zapier). That is one author's experience with a different trigger type, but I think the advice transfers.

Limits and costs to know before you build more

• Claude Code cloud scheduled tasks: the shortest allowed interval is one hour, runs draw down subscription usage the same way interactive sessions do, and Anthropic labels the feature a research preview whose behavior and limits may change.
• Model settings: a scheduled task can be tied to the model you picked when you set it up; Anthropic's cloud task doc says the selected model is used on every run. When a model retires, open your tasks and check that setting. For example, OpenAI's automations page lists GPT-5.5 as retired from Codex with ChatGPT sign-in on October 14, 2026, and asks users to review scheduled tasks in Codex that use it; the page says API-key sessions are not affected.

Event triggers (a task that starts when an email or message arrives) exist in some of these products, but they are a different setup and beyond this guide.

Where to start

Pick the one request you paste into ChatGPT or Claude most often. Decide whether it needs your computer; if it doesn't, use a cloud option. Rewrite it with sources, window, output, success criteria and guardrails, then run that prompt once in a clean chat. Schedule it with the fewest connectors it needs, and read the first two results against the sources before you stop checking. If it survives those two runs, it has earned a place on your schedule.