Claude or ChatGPT · Practical walkthrough

Schedule an AI task and check that it keeps working

The Monday summary arrives every week. Check how you will notice when the source stops returning anything.

Task by Task8 min read

What you'll make

one read-only scheduled task whose every result starts with a status header, a checked first run, and a short note saying who investigates a missing or incomplete run.

What you'll need

Claude (Pro, Max, Ultra, Team or Enterprise) or ChatGPT with scheduled tasks available, a source you’re allowed to use and can also check yourself, and about thirty minutes. A task that sends messages or changes records needs more controls than this guide covers.

Example result.

Status: INCOMPLETE. The Suppliers folder returned zero records; check the source before relying on this result.

Automation doesn’t usually fail loudly; it fails by continuing. The connection loses its sign-in, the folder gets renamed, the source owner leaves, and the task still runs and produces a well-formed summary about nothing.

1. Choose a narrow task and a source you can check

Pick one recurring job, such as a private summary of a supplier folder or a supplier’s public status page. Name the exact source, the period and where the result goes. Avoid a first task that searches the whole organisation for “anything important”.

Run every [day, time and time zone]. Prepare a private summary of new items in [source], covering [window, including time zone]. Deliver it to [destination].

Begin each result with: run time and time zone; requested window; sources accessed, with links; sources missing or only partly retrieved; number of records retrieved, or "count unknown"; and Status: COMPLETED, NO MATCHES or INCOMPLETE.

Use COMPLETED only when the source and window were fully accessible. Use NO MATCHES only when the search completed and found nothing. Use INCOMPLETE if any source, search or part of the window couldn't be checked, or an unexpectedly empty source needs investigation. Never turn an access failure into zero items. Put any gap before the findings. Summarise each change with its date and link, keeping exact figures. Do not send messages, publish, edit records or change files.

Use an explicit time zone and window: every Monday at 08:00 Europe/London, covering the previous Monday 00:00 up to this Monday 00:00. A rolling “last seven days” can silently leave a gap between midnight and the run time.

Why this works: a narrow source is one you can open and compare; a fixed window makes overlap and gaps easier to notice. A missed run still needs recovery.

What you should see: one source, one window with a time zone, one destination.

2. Create the task in your tool

Claude. In Cowork, select Scheduled in the left sidebar, then New task › Create with Claude, answer its questions, review the proposed task and select Schedule. In the newer single Claude experience (no Chat and Cowork options in the message box), describe the task and its frequency in any conversation, then review and Schedule. Scheduled tasks run remotely, so they keep going with your computer asleep; tasks needing local files or apps run only locally. Use a cloud source for this first task.
ChatGPT. Open Scheduled in the sidebar and create the task. Exact delivery times and hourly schedules need an eligible paid plan; Free uses flexible windows (morning, afternoon, night) and recurs at most daily. A task created in a project can’t use that project’s uploaded files. Manage notifications under Settings › Notifications.

Why this works: the scheduling route decides what the run can reach, so a source that works in chat may not work on a schedule.

What you should see: the task listed on the Scheduled page, with the time zone and source you intended. A chat reply saying it will do something isn’t enough.

3. Check the saved instructions

Open the task and compare its instructions with the request in step 1. Verify the exact source, window, destination and read-only scope. The status header must distinguish an incomplete check from a completed search with no matches.

Why this works: a chat confirmation can omit a detail that matters on every later run.

What you should see: the full status header in the saved task, with no added sending or editing step.

4. Check the first result before depending on it

Run it on demand, or wait for the first run. Open the source yourself. Compare the window, at least one known record and every material fact; for a small source, the complete list. Check the notification as well as the stored result: with notifications off, a successful run can look like no run from your inbox.

If the source can’t be reached, check the task’s state and permissions. Don’t grant wider access just to remove a warning.

Why this works: the status header is the task’s own report on itself. Only the first comparison tells you whether that report can be trusted.

What you should see: a header whose sources and count match what you found by hand.

5. Write the missing-run rule

A task can’t send you its status when it hasn’t run. Put an independent reminder in a calendar you already use, early enough to recover before the result matters.

Task and owner:
Backup person or manual route:
Expected result and time zone:
Investigate if no result by:
Investigate if INCOMPLETE, or zero records from a source that is never empty:
Where to check the task and source:
How to pause the task:

Pick the interval from the consequences: a Monday report needed for a Tuesday decision needs a Monday check. The appeal of automation is that you stop thinking about the task. The cost is that you also stop thinking about the task.

Why this works: a missing run and a bad run fail differently, and only an outside reminder catches the first.

What you should see: a filled-in note and a recurring calendar check.

Check the result

The window and time zone in the header match the schedule.
Every source you expected is listed as accessed or missing.
The record count is plausible for that source; zero from a source that’s never empty is a warning, not a quiet week.
A reminder exists outside the tool for a run that never arrives.

The usual failure: an empty or partial run reads like a normal one. Without the header, a renamed folder produces “a quiet week for supplier correspondence”, which is false and unremarkable. Spot it by reading the header’s count and sources first, every time.

The latest result reports [what it reported], but [source or known item] was missing. Revise the task instructions so this case is reported as INCOMPLETE, with the gap before any findings. Keep the schedule and read-only scope. Do not widen permissions. Show the revised instructions before saving.

[what it reported] is the status or summary it gave; [source or known item] is what you found missing.

Worked example: the week the folder was renamed

Fictional example. An operations lead at a wholesaler has a Monday task summarising the past week’s supplier emails from a folder called Suppliers. A normal week:

Run: Mon 7 Sep 2026, 08:00 Europe/London | Window: 31 Aug–7 Sep
Accessed: Suppliers folder | Records retrieved: 6 | Status: COMPLETED
Northgate Packaging — 4% increase on corrugated from 1 November. 2 Sep.
Velmont Foods — October delivery slot moves to the 14th. 3 Sep.

Six weeks later, someone tidies the mailbox and supplier mail starts landing in a folder called Suppliers – 2026:

Run: Mon 19 Oct 2026, 08:00 Europe/London | Window: 12 Oct–19 Oct
Accessed: Suppliers folder | Records retrieved: 0 | Status: NO MATCHES

The failure. Technically accurate: the folder was reachable and empty. An earlier version of the task, without the header, had sent “A quiet week for supplier correspondence.” Quiet weeks happen. Nobody investigates a quiet week.
How it was caught. The window was right, so the schedule was fine. The source was reachable, so it wasn’t a sign-in failure. Zero from a folder that had never returned zero pointed at the source’s contents. The missing-run note’s “zero from a source that’s never empty” line sent the lead to the mailbox.
The repair. She used the repair prompt so that zero records from this folder is reported as INCOMPLETE (the header in the example result), then pointed the task at the renamed folder. An on-demand run returned five records.

The header didn’t detect the problem; a person reading one line did. Its job was to make the problem visible in the two seconds of attention a routine automated result actually gets.

Go deeper

Optional training: Check a known item without treating it as proof

A known recent item can reveal a missing source. Ask for it explicitly, then confirm it yourself.

In the next report, check whether [known item and date] is present in [source]. State found, not found or unable to check, with its link if found. Do not infer complete coverage from finding this item.

Finding it proves only that this item was read. Don’t create fake supplier or customer records just to test a task.

Optional training: Keep ownership and pauses visible

Record the account running the task, its purpose, source, owner and backup, and recheck after a source moves, an account changes or someone edits the prompt. Ownership doesn’t transfer automatically when someone leaves.

Tasks can stop without being deleted: ChatGPT may pause unattended tasks after a period of inactivity, and OpenAI says deleting a task’s chat pauses the task. Check the Scheduled page before tidying old conversations. In Claude, the Scheduled page shows upcoming and past runs, where you can pause, edit or run a task on demand. On Claude Pro, Max and Ultra, tasks created before 6 October 2026 that ran on your computer keep running there, so check which of yours are local.