Claude Scheduled Tasks: Turn Cowork Into a Proactive Teammate

Every module up to here has been about prompts you send. Scheduled tasks flip that. You write the prompt once, pick a cadence, and Cowork runs it on its own — at 6 AM before your day starts, at 11 PM when tokens are cheap, every Friday at 4. This is the feature that turns Cowork from a powerful tool into something closer to a junior teammate.

What a scheduled task actually is

A scheduled task is:

1. A prompt — the thing Cowork will run.
2. A cadence — hourly, daily, weekly, monthly, or one-time.
3. Optional scope — a folder, a project, a model choice.
4. An output destination — a Slack channel, a file, an inbox.

Under the hood, Cowork registers the task with your operating system’s scheduler (cron on Mac, Task Scheduler on Windows). When the trigger fires, Cowork wakes up, runs the prompt with full access to your connectors and skills, and writes the output where you asked.

The practical consequence: your computer has to be on and awake. A sleeping laptop = a skipped run. If you travel with a closed laptop for a week, you will miss runs. Not a dealbreaker — just plan for it.

Where to find it

In Cowork, the left sidebar has a clock icon labeled Scheduled. Click it. You’ll see:

• A list of tasks you’ve created.
• An inbox at the top — runs that completed while you were away, shown like unread messages.
• A New task button.

You can only create scheduled tasks from Cowork, not Chat. That’s because they need local machine access to run.

Building your first task: a daily briefing

The canonical first task. A prompt that pulls tomorrow’s calendar, researches the people you’re meeting, checks past conversations, and pushes a summary to Slack before your day starts.

Click New task. Fill in:

• Name: Daily briefing — Slack
• Prompt:

Go through my calendar for tomorrow and look at
every meeting. Ignore meetings with no external
attendees (kid-related, personal reminders, etc.).

For each remaining meeting:
- The person's name and company
- A one-line description of what the company does,
  from their website
- A link to their LinkedIn profile
- Their most recent 1-2 LinkedIn posts (if any
  from the last couple of months)
- Any transcripts in Granola from past conversations
  with them, summarized in a sentence or two
- Any meaningful email exchanges we've had — ignore
  pure scheduling

Separate each briefing with three dashes.
Push the final output to the Slack channel #test.

• Frequency: Start as Manual. Always. Never schedule a task before you’ve tested it.
• Folder: Leave blank — this task doesn’t touch local files.
• Model: Opus 4.7. This task uses several connectors (Calendar, LinkedIn via Chrome extension, Granola, Gmail, Slack) and benefits from the stronger model.

Save. Then click Run now to test.

What happens. Cowork asks for permissions — Granola access, Chrome extension to visit LinkedIn, Slack write access, Gmail access. Approve each. Let it run. Expect 5–10 minutes on the first execution because it has to warm up every tool.

When it finishes, open Slack. You should see the briefing. Read it. Fix anything wrong in the prompt — “ignore meetings with just me” or “include the company’s latest press release” — and test again.

Then schedule it. Change frequency from Manual to Daily at 11 PM (tokens are cheaper overnight and the briefing is ready when you wake up). Save. You’re done.

Task 2: Automated Twitter research to Notion

A scraper that runs daily, pulls top AI tweets from your Twitter feed, and appends to a Notion database — so you build a curated knowledge base over time without lifting a finger.

Prompt:

Go into my Twitter and find the top AI tweets
by likes, retweets, and bookmarks over the last
24 hours.

Append the results to the Notion database called
"Automated Runs from Cowork." The fields are:
- Tweet text
- Article name (if the tweet links to one)
- URL
- Description
- Username

Make sure you:
- Append new rows, never overwrite
- De-duplicate against tweets already in the database
- Use your best judgment on which fields to populate

Frequency: daily.

What happens. Cowork uses the Chrome extension to open your Twitter account, scrolls and identifies high-engagement AI tweets, cross-references against the existing Notion database, and appends new rows. You end up with a daily-growing knowledge base of what matters in AI without browsing Twitter.

An honest note from people who have built this: you may not need it as much as you think. A lot of scheduled tasks are aspirational — “I’ll read the curated list every morning!” — and in practice nobody reads them. Build for workflows you actually use, not ones you wish you used.

Task 3: Weekly inbox triage (chaining a skill)

Scheduled tasks compose with skills. If you built an inbox-triage skill in the skill-creator module, wrap it in a scheduled task:

Prompt:

Run the inbox-triage skill on my inbox from the
last 24 hours. Send the red and yellow sections
to Slack channel #me. Save the full output to
~/Cowork/inbox-triage/YYYY-MM-DD.md.

Frequency: weekdays at 7 AM.

You wake up to a ranked inbox summary in Slack. The full file stays on your machine for reference. The skill does the judgment; the scheduled task delivers it.

What makes scheduled tasks different

Three things worth internalizing:

1. They run overnight. Token usage is typically cheaper during off-peak hours, and heavier tasks are less disruptive when they’re not competing with your interactive sessions.
2. They compose with connectors and skills. The best scheduled tasks chain three or four tools — Calendar + LinkedIn + Granola + Slack — in a way that would be tedious to do live.
3. They create their own inbox. Runs accumulate in the Scheduled sidebar like unread messages. You check it when you want to, not whenever the task happens to fire.

The traps

• Scheduling before testing. Always run manual first. Always. A task with a typo in the prompt will run every day producing garbage until you notice.
• Closed laptop. You will lose runs when you travel. Either adjust sleep settings or accept it. Don’t schedule mission-critical tasks on scheduled tasks alone — they are not a reliable cron replacement for business-critical workloads.
• Permission dialogs. The first run of any new task will fire permission prompts for connectors. You have to be at the machine to approve them. After the first approval, subsequent runs are silent.
• Cost creep. A task that runs daily with Opus 4.7 and extended thinking across five connectors will burn tokens fast. Check your usage after a week of running, and drop to Sonnet for tasks that don’t need Opus.
• Nobody reads the output. Build tasks that land in a channel you actually check. A daily briefing in an unused Slack channel is landfill.

What to build, in priority order

For most knowledge workers, the highest-ROI scheduled tasks are:

1. Morning briefing. Calendar + LinkedIn + past conversations, pushed to Slack or email before the workday starts.
2. Inbox triage. Wraps your inbox-triage skill. Runs weekday mornings.
3. Weekly digest. A Friday afternoon summary of what you shipped, what’s open, what’s coming. Reads your project files and Granola transcripts.
4. Domain-specific monitor. A daily scan for something narrow — a competitor’s pricing page, a portfolio company’s press releases, a specific Google Alert-style query.

Build the first two. See if you use them. Only add more once you’ve confirmed you actually open the outputs.

Common mistakes

• Starting with a complex task. First task should be simple enough to verify in one manual run. The calendar briefing above is a good target.
• Not watching the first manual run. Permissions prompts appear once. If you miss them, the task hangs.
• Writing a one-shot prompt that assumes state. “Continue where we left off” doesn’t work — each run is a fresh session. Include enough context in the prompt that it can run standalone.
• Using Opus for everything. Most scheduled tasks work on Sonnet. Use Opus only where reasoning complexity demands it.
• Never revisiting the tasks. Scheduled tasks drift. A briefing for people who left the company, a feed from a service you no longer use. Prune every few months.

Where this is heading

Scheduled tasks today are what the course transcript calls “baby agents” — small automated routines that run on a cadence. The direction of travel is obvious: these will become full agents that run continuously, notice when they have work to do without a schedule, and initiate actions. Anthropic is building toward that. Getting comfortable with scheduled tasks now means being ready for the next rung.

Next up

You’ve now covered the complete foundational stack: setup, memory, global instructions, connectors, projects, skills, skill-creator, artifacts, and scheduled tasks. The rest of the Cowork Insider guide moves into specific workflows — the Excel and PowerPoint plugins, the Chrome extension, computer use, and token cost discipline. Start with whichever matches the work you do most. Back to the hub.