A routine answers “when should this run?” A skill describes “how should the work be done?” The official documentation supports scheduled and event-triggered routines. Keep those two decisions separate while designing your workflow.
Prove the one-time task first
Start with a task such as summarizing a specific folder of meeting notes. Ask for a fixed output: decisions, owners, deadlines, and links. Check it with more than one input before giving it a schedule.
Write down what to do when a note has no owner, a date is ambiguous, or two documents disagree. These details determine whether the routine remains useful after you stop watching every step.
Save the method as a skill
Ask the Bot to preserve the sources, selection rules, output format, and approval boundaries. A useful method should be understandable without reading the entire conversation that produced it.
Save the meeting-summary method as a skill. Include the folder to read, the rule for selecting new notes, how to identify decisions, the output columns, and what to report when an owner or date is missing. Preparing a draft does not authorize sending it.Specify the schedule and time zone
Make the intended schedule explicit rather than relying on a relative phrase such as “tomorrow morning.” For example:
Every weekday at 9:00 AM Asia/Shanghai, run the meeting-summary skill on notes added since the last successful run. Return the report in this conversation. If the folder is unavailable, report the failure instead of substituting an old report. Do not contact meeting participants.Inspect the saved schedule. The official guide says a newly created scheduled routine waits for its next scheduled time; use the available test action for an immediate check.
Define a useful result for every run
For a daily digest, “no new notes” is a valid outcome. For a change monitor, you may prefer notifications only when something meaningful changes. Choose deliberately so silence has an understandable meaning.
Keep the source window, deduplication rule, and failure response in the workflow. If you only request “send a report every day,” duplicates and stale results become difficult to diagnose.
Review after the first few runs
Check that the routine used the intended source, produced an actual result, and stopped at the expected boundary. When it does not run, work through our routine troubleshooting checklist before repeatedly recreating it.