Scheduled Tasks & Routines Basics
8 examples to get you started with Scheduled Tasks & Routines - 5 basic and 3 intermediate.
Prerequisites
- A Cowork workspace with at least one connected tool or data source you want a routine to check (a folder, an inbox, a tracker, or similar).
- Know which output you want: a report back to you, a file update, or a message to a channel.
- No Agent Skill is required to start - a plain written prompt is enough for your first routine.
Basic Examples
1. Write the Prompt Before You Schedule Anything
Draft the instruction Claude will follow every time the routine runs, before you touch the schedule settings.
State the task in one or two sentences: what to check, what to produce, and what "done" looks like. For example: "Check the shared status folder for new updates since the last run and summarize them in a short report."
- A routine reruns the same instruction every time, so vague wording produces vague output on a schedule instead of just once.
- Write the prompt as if no one will be there to answer follow-up questions, because no one will be.
- Keep the first version simple - you can add detail once you have seen a few real runs.
Related: How Scheduled Tasks and Routines Work in Cowork - the mental model behind what you are about to set up.
2. Pick a Cadence That Matches How Often the Output Actually Changes
Choose hourly, daily, or weekly based on how fast the underlying information changes, not on how often you want to feel informed.
If the source data updates once a day, an hourly routine will mostly report "nothing new" - which wastes a run and adds noise you have to skim past.
- Daily is the right default for most first routines - status checks, digests, and summaries.
- Hourly suits monitoring where freshness genuinely matters, like watching for new data as it arrives.
- Weekly suits roll-up reporting where a longer view is more useful than a frequent one.
Related: Choosing the Right Cadence: Hourly, Daily, or Weekly Routines - the full trade-offs behind this choice.
3. Scope the Routine's Tool Access Narrowly
Grant the routine access only to the connected tools it needs for this specific task, not everything available in your workspace.
A daily digest routine that only needs to read one folder should not also be authorized to edit files elsewhere - narrower access limits what can go wrong during an unattended run.
- Start narrow; you can widen access later once you trust how the routine behaves.
- Review the access list at setup time, not just the prompt text.
- Unused access is unused risk on a run nobody is watching.
Related: Connected Tools and Skills: What a Routine Can Access While Running - a deeper look at scoping decisions.
4. Save the Routine and Let the First Run Happen on Schedule
Save your prompt and cadence, then wait for the scheduled time rather than forcing an immediate manual run.
Seeing the routine trigger on its own is the real test - it confirms the schedule is wired up correctly and that Claude can complete the task without you present to prompt it.
- Note the exact time you expect the first run so you know when to check back.
- Resist the urge to babysit the first run in real time; the point is to confirm it works unattended.
- If your platform offers a manual "run now" option for testing, use it once, then let the schedule take over.
5. Read the Output as Carefully as You Would a Person's First Report
Treat the first scheduled output like you would a new team member's first status report - read it fully, not just the headline.
Check that it actually reflects what happened, that it did not silently skip a step, and that the format is something you would want to receive again tomorrow or next week.
- Compare the output against what you know actually changed, if you can.
- Look for signs the routine ran out of things to say versus genuinely found nothing new.
- If anything reads oddly, that is signal to revise the prompt before the next run, not to ignore it.
Intermediate Examples
6. Tighten the Prompt Based on What the First Few Runs Actually Produce
Revise the original prompt using patterns you noticed across two or three real runs rather than guessing at improvements in advance.
If the output keeps missing a category of update, or including something you don't care about, add a specific line to the prompt addressing exactly that - specific fixes work better than a general rewrite.
- Change one thing at a time so you can tell what improved the output.
- Keep a short note of what you changed and why, since you will forget by next month.
- Expect two or three rounds of tightening before a routine feels reliably "done."
7. Confirm the Routine Handles a Run With Nothing New to Report
Deliberately check what the routine produces on a day when the underlying source has no changes.
A well-written routine should say clearly that there was nothing new, rather than inventing filler content to look productive - this matters because you will not be there to catch a fabricated update in real time.
- A short, honest "no updates" report is a sign the routine is well-scoped.
- A padded-out report on a quiet day is a sign the prompt needs a clearer instruction to report absence plainly.
- This check matters more for hourly and daily routines, where quiet periods are common.
Related: What Happens When a Scheduled Task Runs Without You Watching - what else to expect from unsupervised runs, including failures.
8. Graduate a Reliable Prompt Into a Packaged Skill
Once a routine has proven itself over several runs, consider packaging its instructions as an Agent Skill instead of leaving it as a standalone prompt.
This is worth the extra setup when the routine's logic is something you want to reuse elsewhere, share with teammates, or protect from accidental edits to the underlying instructions.
- Not every routine needs this step - plenty of routines stay as plain prompts indefinitely.
- A Skill is worth it when the same logic is valuable outside this one routine.
- This is a natural next step once you are comfortable with the basics covered above.
Related: Setting Up a Weekly Digest Routine Step by Step - a full worked example of a more advanced routine.
Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.