Scheduling and automation

Every lesson so far has had you at the keyboard: you open a chat, you ask, Claude answers. Useful, but it only happens when you show up. This lesson is about the opposite. A scheduled task is work you describe once and then never start again. It runs on its own, on the cadence you set, and hands you the finished thing: a brief in your inbox each morning, a report every Monday, a watch that pings you when something changes.

If you have done any of Part 5, you have met Cowork, the mode where Claude does multi-step jobs with your files and connected tools. Scheduled tasks are Cowork with a clock on it. Same powers, minus the part where you have to be there.

Last checked: July 2026. Scheduled tasks are in beta and rolling out across plans. Cadences, buttons, and limits may shift, so treat the specifics here as a snapshot and check the linked docs for the current detail.

What a scheduled task actually is

You describe the job once. Claude saves your prompt as the task's instructions and runs them on your schedule, or on demand when you press the button. Each run happens in its own Cowork session, so you can open it later and read exactly what it did.

The key word is finished. A scheduled task does not remind you to write the report. It writes the report. You come back to output, not to a nudge.

A scheduled task has the same reach as any Cowork task. It can search your Slack, read your files, run web research, pull from a connected tool, and build a report, using whatever connectors, skills, and plugins you have set up. Scheduling does not give Claude less to work with. It just decides when.

Where it lives, and what it costs

Scheduled tasks are part of Cowork, so they live on paid plans only: Pro, Max, Team, and Enterprise. You will not find them on Free. On Team and Enterprise an admin controls Cowork access, so if you do not see it, that is the first thing to check.

You manage everything from one place: the "Scheduled" item in the left sidebar. That page lists every task you have made, shows upcoming and past runs, and lets you open any single run to see what happened. There is nothing extra to install and no separate bill; a scheduled task draws on the same usage as the rest of your Claude account.

Cloud or local: where the task runs

This is the one detail people trip on, so it is worth getting straight. Most scheduled tasks run in the cloud. They use your connectors and the files saved to your Claude account, and because they run remotely, they fire on their cadence even when your computer is asleep or the desktop app is closed. That is the whole appeal: the morning brief lands before you open your laptop.

There is a catch, and it is a fair one. If a task needs files or apps on your own computer, it has to run locally, which means your machine must be awake and the app open when the clock strikes. Cloud runs cannot reach into a folder on your hard drive; they work with what is in your Claude account and your connectors.

The rule: connected tools and account files, it runs in the cloud, any time. Local files or local apps, it runs on your machine, so keep it awake.

Two ways to make one

You can build a scheduled task by talking, or by filling in a form. Both start from the "Scheduled" sidebar and the "New task" button.

Create with Claude is the conversational route, and the easy one. Pick it and Claude opens a task pre-loaded to help you set the schedule up. It may ask a few multiple-choice questions to pin down the detail, then it shows you the plan: the task's name, the cadence it will follow, and what it will actually do. You confirm by clicking "Schedule", and it is live. If you are already inside a Cowork task, you can trigger the same flow by typing /schedule.

Set up manually is the form. You fill in the fields yourself:

Task name:      a name you'll recognise later
Prompt:         exactly what the task should do (the important one)
Approval mode:  whether Claude pauses for your sign-off on risky steps
Cadence:        hourly, daily, weekly, weekdays, or manually
Model:          optional; leave the default unless you have a reason
Folder:         optional; only if it needs a specific place to work

Manual is quicker once you know what you want. Conversational is better when you do not, because Claude helps you shape the job. Use whichever fits.

The one habit that saves you: run it once first

Here is the mistake to avoid. You dream up a clever automation, schedule it daily, and a week later you have seven mediocre reports and no idea which instruction went wrong.

Do this instead: run the job once as a normal Cowork task, by hand. Read the output. Is it the right length? The right source? Does it skip the noise and keep the signal? Fix the prompt until one run is genuinely good. Then schedule it. A scheduled task is only ever as good as the prompt behind it, and a schedule just multiplies whatever that prompt produces, warts and all.

Writing a prompt that holds up alone

A chat prompt can be loose, because you are right there to steer. A scheduled prompt cannot, because no one is there to answer a follow-up. When the task runs at 6am while you sleep, Claude cannot ask "which report did you mean?" So say it all up front. A good scheduled prompt names four things:

• The source. Where the material comes from. "My Google Calendar and unread Gmail from the last 24 hours." Vague source, vague result.
• The output. The exact shape you want back. "A summary under 200 words, with a bullet list of today's meetings and a short 'needs a reply' section."
• What to skip. The noise to leave out. "Ignore newsletters and automated receipts." This is what turns a wall of text into something you actually read.
• The edge cases. What to do when there is nothing. "If there is no urgent email, just say so in one line."

Here is the difference in practice:

Weak:   Send me a summary of my emails.
Strong: Each weekday at 7am, read my unread Gmail from the last 24 hours. Write a summary under 150 words. List anything that needs a reply as bullets, newest first. Ignore newsletters, receipts, and calendar invites. If nothing needs a reply, say "Inbox is clear" and stop.

The strong version runs unattended for months without you touching it. The weak one needs a conversation every morning, which defeats the point.

Good first automations

Do not start with the moon landing. Start with something small, repetitive, and low-risk, and build trust from there. Strong opening jobs:

• A daily briefing. Your calendar, unread email, and overnight Slack, boiled down to one short read before the day starts.
• A weekly report. Numbers from a connected sheet or tool, pulled into the same format every Monday, so you stop rebuilding it by hand.
• A recurring watch. Track a competitor, a topic, or a bit of industry news, and get a short digest on a set cadence.
• A tidy-up. Sort, rename, or process files in one folder on a schedule so it never becomes a swamp.
• A standup summary. Turn yesterday's activity in your project tool into a plain "here's where things stand" note for the team.

Notice the shared shape: a clear source, a clear output, and no risky action. Those three make a job safe to hand over. The best automations are the boring, repeating chores you would never miss doing, which is exactly why they are worth giving away.

Managing your tasks

Once a task is live, the "Scheduled" page is mission control. From there you can review upcoming and past runs, open any run to read what it did, edit the instructions or the cadence when your needs change, pause a task and resume it later, run it on demand without waiting for the clock, and delete it when it has served its purpose.

Two habits keep a stack of automations healthy. Check the early runs. For the first week, actually read the output and make sure it is doing what you meant, then trust it. And prune. A task you set up for a project that ended is just noise now. If you are not reading a report, pause it or bin it.

Staying in control

Automation is leverage, and leverage cuts both ways, so a few guardrails.

• Keep a human on anything that acts. A task that reads and reports is safe to leave alone. A task that sends, deletes, or spends deserves the approval mode, so Claude pauses for your sign-off before the risky step. Reporting on autopilot is great; irreversible actions on autopilot are how you get a nasty surprise.
• Mind what it can reach. A scheduled task inherits your connectors, which means it can touch whatever they touch, on its own, while you are away. Grant it the access the job needs and no more, and glance at the safety guidance before you wire it into anything sensitive.
• Automate the chore, not the judgement. These tasks are brilliant at the repetitive gathering and shaping that eats your time. The decision about what the output means, and what to do next, is still yours. Let the task lay the table; you still choose the meal.