Try safe automation with a manual fallback

Use this task: “Remind only me to review my fictional Oakfield invitation.” The input is course practice text; the result is a private reminder. It may not send the invitation, make a booking, purchase anything, delete data or access other accounts.

A useful automation specification names the trigger, input, action, destination, review and stopping point. For this test: one chosen time, the reminder sentence, a reminder to you, a check of its actual result, then no further runs.

Run the prompt below in an ordinary chat, or make the checklist yourself. It explicitly requests a manual rehearsal and does not create a schedule. Expected checklist: date/time, venue, free cost, booking deadline, notebook/laptop, and unsupported claims.

Apply the checklist once to your invitation. If a date is missing or sources conflict, stop and flag the gap. Do not invent a date merely to let a workflow continue.

Optional scheduling route: only if your account already supports it, use its documented scheduling feature to request one private reminder about five minutes from now: “Review the fictional Oakfield invitation against the course fact sheet.” Specify today’s actual date, a specific time, your actual time zone and one occurrence. Do not use the fictional 2030 event date as the reminder date.

For example, in a supported ChatGPT account you can ask in chat for this one-time reminder, then open Scheduled to inspect the saved task. If that feature is absent, follow the manual route rather than assuming the chat created a timer.

Before confirming, inspect the destination, date, time zone, exact text and one-time frequency. Use only an existing private notification destination. Do not enable new app connections, expand permissions, buy a plan or set up email sending for this exercise.

Manual route: write the same reminder in your notes and run the checklist now. An existing calendar reminder can prompt you to do this later if you choose. If only the reminder is automatic, describe it as an automated reminder with a manual AI task, not an automated AI workflow. Claude schedules are documented for paid plans; ChatGPT and Gemini access also depends on account and feature.

If you scheduled a test, open the actual saved schedule/task list and verify it exists with the right settings. Then check the first notification or run result. Read the result itself: a task marked “completed” can still contain a bad answer. Find the pause/delete control and cancel any unintended repeat.

A chat promise alone is unverified. A saved schedule is “scheduled”. An observed result is “ran”; actual receipt is “received”. If no result arrives during your practice, record “scheduled; delivery not yet verified”, check status and time zone, and use the manual checklist. Do not claim success or keep creating duplicates.

Record: route used; trigger/date/time zone if applicable; destination; saved schedule checked yes/no; result observed yes/no; review completed yes/no; and how to stop it. For the manual route, write “Manual rehearsal; no AI schedule created.”

Keep the manual checklist as the fallback if a service, notification or permission fails. This is a low-risk learning test, not a suitable way to guarantee exact-time medical, financial or other critical alerts. A paid or missing feature does not prevent course completion.