How Scheduled Tasks and Routines Work in Cowork
A routine is a way of telling Claude to do something more than once without you asking each time.
//
Search across all documentation pages
A routine is a way of telling Claude to do something more than once without you asking each time.
You write the instructions once - a prompt, or a packaged Agent Skill - and attach a schedule to it.
From then on, Claude runs that instruction on its own, at the cadence you picked, with the same access to connected tools and Skills it would have if you had triggered the work yourself.
This page builds the mental model for how that works before the later pages in this section cover cadence choices, failure handling, and setup steps in more depth.
At its core, a routine has two parts: the instruction and the schedule.
The instruction is either a plain prompt - a written description of what to do - or a packaged Agent Skill that bundles instructions with any supporting files or steps it needs.
The schedule is the cadence: hourly, daily, or weekly.
Once both are set, Cowork takes over the job of starting the run.
You do not open a chat and ask Claude to check something; Claude checks it on its own, at the time you configured.
Think of a routine as a standing appointment on Claude's calendar rather than a one-off conversation.
The appointment always has the same agenda - the prompt or Skill you packaged - but Claude runs the agenda fresh each time, using whatever it finds at that moment.
When a scheduled run starts, Claude begins the same way it would if you had typed the request yourself.
It reads the packaged prompt or Skill, decides which connected tools it needs, and works through the task.
The difference is that no one is present to answer follow-up questions, approve an ambiguous step, or redirect Claude mid-task.
That absence shapes how a routine has to be written: instructions need to be specific enough that Claude can complete the work without back-and-forth clarification.
A routine's access to connected tools and Skills is fixed by whatever was authorized when the routine was set up.
It does not gain new permissions mid-run, and it does not lose them either - the scope stays constant for the life of that scheduled run.
This is the same scoping model that governs any autonomous Cowork task, scheduled or not; a routine is simply that same autonomous behavior repeated on a timer.
Trigger (time or cadence hits)
-> Claude loads the packaged prompt or Skill
-> Claude works the task using its authorized connected tools
-> Claude produces output (report, update, file change, message)
-> Run ends; the routine waits for its next scheduled triggerBecause there is no live supervision, a well-built routine also needs to account for what happens when something goes wrong partway through - a connected tool is unavailable, a data source returns nothing, or a step fails outright. That handling is covered in depth on the unattended-run page linked below.
Routines scale well for work that is genuinely repetitive and well-defined, but they are a poor fit for anything that requires judgment calls only a person can make in the moment.
A routine that summarizes new support tickets every morning is a strong candidate; a routine that decides whether to escalate an ambiguous customer complaint is not, unless the escalation criteria are spelled out clearly enough that Claude can apply them without asking.
As routines accumulate, the practical concerns shift from "does this work" to "does this still make sense to run" - stale routines that no longer reflect current priorities are easy to forget about precisely because they run unattended.
| Approach | Strength | Weakness | Best Fit |
|---|---|---|---|
| Manual, on-demand request | Full context and judgment applied every time | Only happens if someone remembers to ask | One-off or unpredictable work |
| Scheduled routine (prompt-based) | Consistent, no reliance on memory | Instructions must anticipate variation without a person present | Recurring reports, monitoring, digests |
| Scheduled routine (Skill-based) | Reusable, packaged logic shared across runs or teams | More setup investment up front | Recurring work with steady structure or steps you reuse elsewhere |
Yes. Each scheduled run works from the packaged instructions fresh - it is not a continuation of a previous chat, though the routine can be told to reference prior outputs if the task calls for it.
No one is watching in real time during a scheduled run, so a routine should be written with enough detail that Claude does not need to pause for clarification. Ambiguous instructions are one of the more common causes of unsatisfying unattended output.
Not inherently - a well-written prompt can handle plenty of recurring work. A Skill becomes worthwhile when the task has enough structure or reusable steps that packaging it formally saves effort over rewriting the prompt each time.
You choose the cadence when you set the routine up - hourly, daily, or weekly - based on how fresh the output needs to be. See the cadence page linked below for the full trade-offs.
Only if the prompt or Skill is written to reference previous output, such as a data source, file, or log that carries state between runs. A routine itself does not automatically retain memory of prior runs.
Future scheduled runs use whatever access is currently authorized at the time they start; a revoked tool will no longer be available on the next run.
The cadence options are hourly, daily, or weekly; pick the closest fit and adjust the prompt's scope if you need finer control over frequency within that window.
A routine is a Cowork task with a schedule attached. Everything true of an autonomous Cowork task - tool scoping, unattended execution - applies to a routine as well.
A single daily prompt that checks one data source and reports back is enough to be useful, and it is the recommended starting point before adding cadence variety or broader tool access.
No - a plain prompt is enough to schedule a routine. A Skill is an optional, more structured way to package the work when it is worth reusing.
Start with the basics walkthrough linked below, which sets up a first daily or weekly routine and confirms it runs unattended before you add complexity.
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.
Reviewed by Chris St. John·Last updated Jul 19, 2026