How to schedule a Claude skill

Bottom line: you cannot schedule a skill from Claude chat — scheduling lives in Cowork. And the limitation that catches teams out is that a Cowork scheduled task runs on your own machine: if your laptop is closed or the desktop app is shut, it does not run. That makes it fine for personal routines and risky for anything a team depends on.

We hit this live on our August 2026 Claude Skills for Finance webinar: asked to schedule a daily ARR summary, Claude declined and explained there is no scheduling tool in chat. The fix took thirty seconds once you know where to look.

Can you schedule a Claude skill?

Yes, in Cowork rather than chat. Chat and Cowork have converged enough to be confusing, so the practical distinction is worth holding onto: chat is where you do ad-hoc work and build the skill; Cowork behaves more like a coworker, handles longer-running tasks, can operate a browser, and is the surface that owns scheduling.

Anthropic's getting started with Claude Cowork covers the surface itself. Note that Cowork availability depends on your plan, so check before you design a routine around it.

How to schedule one

1. Build and save the skill in chat first. Get the output right by prompting, then package it. Scheduling a skill you have not validated just automates a bad answer.
2. Open Cowork rather than a chat window.
3. Ask for the schedule in plain language — name the skill, the time and the timezone. "Schedule the ARR-by-close-date skill to run at 9am PT every day" is enough.
4. Say where the output should land. A summary posted into a Slack channel is more useful than one sitting in an app you have to open.
5. Verify the first real run rather than assuming it fired.

The limitation nobody mentions

A Cowork scheduled task runs through your own Claude instance. Close the laptop and nothing happens; the task waits until the app is open again. For a personal morning summary that is usually harmless — it arrives late rather than never.

Where it becomes a real problem is dependency. If someone on your team owns a scheduled skill that the rest of you rely on and they go on holiday, that output silently stops. Nobody gets an error; the report just is not there. This is also why you saw a run of posts about leaving a Mac mini switched on: it is a genuine workaround, and it is a sign the capability is not yet built for team-critical work.

The rule of thumb: schedule personal routines in Cowork, and put anything the team depends on somewhere that runs server-side.

A worked example: a daily ARR note in Slack

The most common first scheduled skill in a finance team is a short daily or weekly movement note. Here is the shape of it.

1. Get the query right in chat first. Ask for new ARR by month. Then correct it — almost every contract table has several date fields, and the agent will pick one without telling you which. Specify close date rather than start date, or whichever your team reports on.
2. Save that as a skill once the output matches what you would have built by hand. The skill now carries the date-field decision, so you never re-explain it.
3. Move to Cowork and ask for the schedule, naming the skill, the time and the timezone.
4. Route the output to Slack rather than leaving it in the app. A note nobody opens is not a report.
5. Watch the first two real runs. The first proves it fires; the second proves it fires when you were not watching.

Note what carried the value there: not the scheduling, but the date-field decision baked into the skill. Scheduling an ambiguous query just delivers the same ambiguity on a timetable.

What happens when a scheduled run fails

Usually nothing visible, and that is the problem. A missed run does not announce itself — you get silence where a message should have been, which is far easier to miss than an error.

Two habits make this survivable. First, have the skill state its own inputs in the output: the period it read, the row count, the timestamp. A note that says which data it used is one you can sanity-check in three seconds. Second, decide who notices. If the answer is nobody, the routine is decorative, and you should either give it an owner or stop pretending it is part of the process.

For anything genuinely load-bearing, the honest conclusion is that a laptop-dependent scheduler is the wrong home for it, and no amount of care in the skill fixes that.

Choosing when it fires

Cadence design is where most scheduled finance routines go wrong, and the fix is to work backwards from your close calendar rather than from convenience.

Daily works for anything reading a live source that does not depend on close — pipeline, bookings, bank balance, ARR movement. Fire it early enough to be read with coffee.

Weekly suits a cash position or a pipeline-to-plan view. Monday morning is the obvious slot and usually the right one.

Monthly is the one to be careful with. A routine that fires on the third against a ledger that closes on the fifth will be wrong every month. Schedule from close plus a buffer, not from the calendar, and have the skill verify the period is closed before it writes anything — the same discipline we describe in FP&A implementation steps.

Where scheduling actually belongs

The trade-off is about who is depending on the output, not about sophistication.

Platform-native scheduling is the durable answer for recurring finance work, because it does not depend on a person's hardware being awake. More on structuring that in automating FP&A workflows.

What to schedule, and what not to

Good candidates share a shape: the inputs arrive on a predictable cadence, the output is a summary rather than a decision, and being a few hours late is survivable. A daily ARR movement note, a Monday cash position, a flash revenue view a few days after close.

Poor candidates are the mirror image. Anything feeding a board pack, anything where a silent failure would be discovered by someone senior, and anything whose inputs are not reliably closed at the moment it fires. A skill that runs on the third of the month against a ledger that closes on the fifth will confidently produce the wrong number every single time.

Build the check into the skill as well — the same instinct as a tie-out tab in a model. Have it verify record counts, or confirm the period it read is actually closed, and say so in the output. See AI accuracy and auditability for how to structure that.

Scheduling in Aleph

For finance workflows that a team depends on, scheduling belongs next to the data rather than on an analyst's laptop. Aleph runs agent workflows against live ERP and CRM data with each user's own permissions, so a recurring flash or variance summary reads from a governed layer and drills back to source.

See the Aleph Agent, watch the scheduling walkthrough in the webinar recording, or start from the Claude skills for finance guide.