Specify the time, input, output, and notification rules for one recurring task.
A concrete example
Example: at 7:00 AM Asia/Tokyo each day, prepare up to eight new N1 questions based on yesterday’s mistakes. If there were no answers, inspect older records and say why.
What a good result looks like
A run should report the date and answer totals it read, why it chose a question type, and where the draft was actually saved. It should report missing evidence or write failures.
A prompt you can use
Save a recurring task for every day at 7:00 AM Asia/Tokyo. Using my authorized JLPT Master plugin, read the previous day’s N1 answers; if there are none, inspect older history and state the basis. Write up to eight new questions of the best-supported weak type, each with an answer and explanation for every option. Save them as a practice draft, not a formal practice. Reduce the count when evidence is insufficient; never invent statistics or textbook sources. Check that the draft was saved and leave its title and location in the task result. Notify me in ChatGPT only if authorization expires, saving fails, or I need to decide something. Confirm the saved schedule, time zone, and next run.
Three practical steps
First test a read-only request and a manual draft save with the required permissions.
Ask the dot to save the schedule. Check its time zone and next run in Scheduled.
After the first run, inspect Activity and the actual draft, including answers, explanations, and sources. Change or disable the schedule in Scheduled.
Where to check the result
Open Practice Drafts and review the saved questions. A draft is separate from formal daily practice.
View practice drafts →What to watch for
A schedule attempts the saved workflow at its configured time; it does not by itself maintain an evolving learning goal. A delegated local Codex task needs your computer online with the ChatGPT desktop app open. A completed run alone does not prove the draft was saved.
JLPT Master