Most write-ups of Claude scheduled prompts stop at "set a cadence, get a digest." Pick daily or weekly, write a prompt, Claude runs it and drops a summary in your inbox. True, and useful for a morning email briefing or a file-cleanup chore. It's also the least interesting thing you can do with one.
The direct answer: Claude scheduled prompts let you save an instruction once and have Claude re-run it automatically — hourly, daily, weekly, weekdays only, or on demand — without you opening a chat window. The mechanism is simple. What most guides skip is the part that makes it more than a fancy reminder: what the AI is allowed to touch between runs, what happens when it isn't sure, and whether your laptop even needs to be on. Here's the fuller picture, including the honest limits.
What are Claude scheduled prompts, exactly?
A scheduled prompt is an instruction Claude saves and re-runs on a cadence you pick, instead of one you have to retype every time. In Claude Cowork, you name the task, write the prompt in plain language, and choose how often it fires: hourly, daily, weekly, weekdays only, or manually. Claude confirms the schedule and gives the task an ID you can reference later.
Anthropic uses "scheduled tasks" and "scheduled prompts" more or less interchangeably in its own materials, and so does most of the web — if you searched one term, the results for the other apply too.
Worth a quick aside if you're a developer rather than a PM: Claude Code (the CLI) has its own, separate scheduling primitives — /loop for session-scoped polling, and cloud-based Routines for automation that survives independent of any machine. Those are built for babysitting a pull request or a CI run. This guide is about the everyday, non-technical version: Cowork scheduled tasks that check on your actual work.
How do you actually set one up?
In Claude Cowork, click Scheduled in the sidebar, then New task, then Create with Claude. You describe what you want in a sentence — "every weekday morning, check my inbox and draft replies to anything urgent" — and Claude fills in the name, the prompt, and asks you to confirm the cadence.
You don't need to remember task IDs to manage things day to day: ask Claude in plain language, "what scheduled tasks do I have?" or "cancel the deploy check job," and it handles the lookup. Under the hood each task keeps a short ID, and an account can run up to 50 at once — more headroom than almost anyone needs before they start being deliberate about which ones actually earn a slot.
Do you need to leave your computer on?
The honest answer is more conditional than most quick explanations let on, and getting it right changes how you actually use the feature. Claude Cowork scheduled tasks run in the cloud by default — the task fires on schedule even if your computer is off, asleep, or the desktop app is closed. The one exception: tie a task to a local folder so it can read or write real files on your machine, and that task needs your computer on and the app running, because Anthropic's servers can't reach your local filesystem on their own.
So the real question isn't "local or cloud" as a fixed property of the whole feature — it's "does this particular task touch local files?" A daily board sweep that only needs a connector (an MCP-connected project tool, say) can run entirely in the cloud, laptop closed. A task that needs to read files sitting on your desktop cannot. Get that distinction backwards and you'll either keep your laptop on for a task that never needed it, or get caught out by a task that quietly stopped firing because it was tied to a local folder you'd forgotten about.
For automation that needs to survive even longer or reach more infrastructure than a Cowork task's session model, Anthropic's separate Routines feature also runs on Anthropic's cloud infrastructure — with its own tradeoff of working from a fresh, git-committed snapshot rather than live local files.
What should you actually schedule?
The examples that circulate most — a morning email digest, tidying up files, a Stripe or invoice summary — are real and fine, but they're also the easy 20%. They're read-and-summarize jobs: Claude looks at something, tells you about it, and nothing changes until you act on the summary yourself.
The more useful case is pointing a scheduled prompt at something Claude can actually work, not just describe — a shared project board is the clean example. Connect Claude to a board over MCP once, then schedule a daily prompt along the lines of: "check the board, triage anything new, move stale cards, draft cards from what's in the notes, and give me a short summary." Instead of a passive digest, each run does real work and only the parts that need your judgment come back to you.
We wrote the fuller version of that specific workflow — what to hand it first, how to keep it from making mistakes — in how to use Claude for project management. This post is about the scheduling mechanism underneath it; that one is about the habit you build on top of it.
What happens between runs?
This is the honest question competitor write-ups skip, because most of their examples never touch anything a second person also depends on. A daily email digest can't step on anyone's toes. A scheduled prompt that moves cards, reassigns work, or closes tickets on a shared board can — if you're not careful about what it's allowed to decide on its own.
Two things matter here. First, escalation instead of blind action: a well-set-up scheduled prompt shouldn't close ambiguous cards or reassign someone else's work quietly. In Comuna, anything the AI isn't confident about opens a small approval request instead — you approve, ask for changes, or reject, and the run continues from your answer next time. Second, attribution: every action a scheduled run takes should carry the AI's identity, not get folded silently into your account, so six months from now you can still tell what a human did and what ran on autopilot at 6am.
Be plain about the "pull, not push" part, because it's the thing this whole category quietly implies away. A scheduled prompt doesn't watch anything continuously and react the moment something happens — it wakes up on the cadence you set, does its pass, and goes back to sleep. If something urgent lands ten minutes after a run finishes, nothing reacts to it until the next scheduled fire. That's a fine tradeoff for a daily sweep; it's the wrong tool if you actually need real-time reaction to an event as it happens.
Where scheduled prompts don't fit
Two honest limits, beyond the local-folder dependency above. If you need something to react the instant an event occurs — a customer reply, a failed deploy, a support ticket landing — a fixed cadence is the wrong shape; look for an event-driven trigger instead of a timer. And if the task involves an irreversible action on shared work (deleting something, notifying a whole team), don't hand a scheduled prompt full autonomy over it without an approval step — build in the escalation, or keep that one step manual.
Are Claude scheduled prompts the same on every plan?
No. Cowork scheduled tasks are available on a paid Claude plan (Pro and up) rather than the free tier, and Max plans get more usage headroom on top of the same feature. Confirm current plan details on Anthropic's side, since pricing and limits change independently of this post.
How often can a scheduled prompt run?
Cowork gives you hourly, daily, weekly, weekdays-only, or manual cadences — you don't write cron syntax, you describe the cadence and Claude sets it. (Developer-facing tools like Claude Code's /loop do support finer-grained intervals, but that's a different feature aimed at a different job.)
Can a scheduled prompt work on a real project board, not just send me a summary?
Yes, if the AI is connected to the tool via MCP with write access — it can then create, move, and comment on real cards on each run, not just report on them. See how to use Claude for project management for the full workflow, and connect Claude in 60 seconds for the setup steps.
What if I'm often away from my computer?
Only tasks tied to a local folder need your machine on and awake — keep a task cloud-only (connectors, no local files) and it fires whether your computer is on or not. If a task genuinely needs local files and you're away a lot, that specific automation is a better fit for Routines instead.
Comuna is free forever — no credit card, bring your own AI. Spin up a workspace and try it.