How to automate recurring work with Claude: a practical guide

To automate work with Claude, choose one repeated task, define a correct result, and test it by hand before scheduling it. Use Claude for the steps that need judgment, and use scripts or app integrations for predictable data movement.

Start with the chore you are tired of doing by hand every week.

What is in this guide?

• Pick a useful first task
• Write the procedure
• Choose where it runs
• Build a weekly brief
• Test failures and measure value
• Common questions

What should you automate first?

Look for work that repeats, uses accessible inputs, and produces a result you can check. Start with preparation: summaries, draft reports, sorting, or a list of exceptions to review.

Ask Claude to interview you about one task:

What starts this task? Which inputs do I use? What decisions do I make? What does a correct result look like? What should happen when information is missing?

Record the time you spend doing it now. Include cleanup and checking. Pick a task where a small failure is easy to notice and undo.

Avoid starting with “manage my business.” It hides too many decisions. “Prepare a draft Monday status report from these three sources” is a job you can test.

How do you turn a task into a reliable procedure?

Describe a completed example. You can type it out, dictate a voice memo, or walk through the steps in a supported tool, then ask Claude to turn that example into a repeatable procedure.

Don’t assume a single demonstration captures all the rules. The agent still needs to know how to handle late data, duplicate inputs, and exceptions.

Use this template:

Job: [one recurring result]
Trigger: [schedule or event, with time zone]
Inputs: [approved sources and how fresh they must be]
Steps: [the normal path]
Decisions: [where interpretation is needed]
Output: [file, format, and destination]
Missing data: [stop, skip, or report a partial result]
Permissions: [what may change; what needs approval]
Duplicates: [how to detect work already done]
Validation: [the checks that prove the result is usable]
Failure report: [where to report a problem]
Owner: [who maintains the workflow]

When a procedure stabilizes, save it as a reusable skill. In Claude Code, that means a SKILL.md file inside a skill folder. The skill supplies the instructions; the scheduler supplies the trigger. Claude Code skills.

Keep the procedure and its source data separate. A skill explains how to prepare a brief; it shouldn’t store last week’s private customer records.

For a larger workflow, map the chain: input → operation → output → owner. Mapping makes shared dependencies obvious before two automations start overwriting the same records.

Which automation option should you use?

Choose the runtime based on where the inputs live and when the work must run.

Claude documents these scheduling choices in its common workflows. For the other options, see Zapier automation, Make, and Hermes documentation.

Anthropic announced a unified Claude experience on September 16, 2026. It brings former Cowork features into ordinary conversations, with rollout starting on Pro and Max. Its examples include recurring reports. If you prefer the app, check which task and scheduling controls your account has before choosing a coding-oriented setup. Announcement and rollout details.

When is a loop enough?

Use a loop for a temporary check, such as watching a build during a working session. The current CLI documentation says recurring tasks expire after seven days. They depend on the session lifecycle. Use a durable scheduling option for an ongoing business job. Claude Code scheduled prompts.

When should the task run locally?

Use Desktop scheduling when the task needs local files or tools. Confirm the computer can stay awake and the app can run. Also check whether the task uses the current working directory or an isolated worktree. Uncommitted files can matter. Desktop scheduled tasks.

When should the task run in the cloud?

Use a cloud routine when work must continue with your laptop closed. Configure its environment, repositories or data sources, network access, and credentials explicitly—a file sitting on your laptop isn’t automatically available in that cloud environment. Claude Code routines.

When does the task need no model?

Use ordinary code for stable rules. A script can validate a date, normalize a URL, or calculate a total. Bring in the model only where the step requires judgment, like grouping related issues or drafting a clear summary.

Scripts still take time to write and maintain, and hosting them can cost money too. The advantage is that predictable operations don’t need a fresh model decision on every run.

Example: build a weekly project brief

Suppose you want one Monday brief from a project tracker and a folder of meeting notes. The result should list completed work, blocked work, and decisions you need to make.

1. Run it by hand

Give Claude a small sample and this request:

Prepare a weekly project brief from these sources. Link each claim to its source. Keep completed work, blockers, and decisions separate. Flag stale or conflicting information. Save a draft. Do not update the tracker or contact anyone.

Review the draft. Correct the structure and the decision rules. Save one good result as an example.

2. Separate the data work from the writing

Have a connector or script collect the relevant records. Limit the date range. Include record IDs, source links, and update times.

Then let Claude write from that bounded input. Do not ask it to browse the entire company archive each week when only twenty records changed.

3. Test the awkward cases

4. Add the schedule

Set the owner, time zone, output location, and failure alert. Run a test with safe inputs and read the output yourself—don’t assume “scheduled” means “working.”

Keep no-change runs quiet, and notify the owner only when an input fails or a decision needs attention. Put that notification rule right in the task definition.

5. Review the first week

Check the output against the original sources, counting any corrections or missed items. Keep outbound messages and source-record edits out of scope until the draft workflow proves reliable.

For repository-based automation, Anthropic provides a GitHub Actions integration. Follow that documented setup instead of inventing a custom workflow schema.

How do you know the automation helps?

Measure time saved after review, fixes, and maintenance. A long report nobody reads is not a win.

Here is an illustrative calculation:

These numbers are an example, not a claim about Claude's performance. Track your own runs over several weeks, and include software and model costs when deciding whether the automation is worth keeping.

Watch for silent failures: stale credentials, renamed fields, a sleeping laptop, or an empty input mistaken for “nothing happened.” A reliable automation tells you when it failed and shows where it stopped.

How should the system grow?

Keep a short list of working jobs, their owners, and their dependencies. Add the next task after the first one survives ordinary errors. Remove an automation when its maintenance costs more than the work it saves.

Miles Deutscher's automation thread inspired this guide. The steps are my practical interpretation; product details come from the linked official documentation.

For setup, use the Claude Code field guide. For reusable procedures, use the agent skills guide. If the goal is a public-facing chat assistant, start with the website assistant guide.

FAQ

Can I automate work with Claude without coding?

Yes. You can describe a task, use supported connectors, and configure an available scheduling feature. Complex integrations may still need code or technical help.

Is a Claude skill an automation?

A skill is a procedure. It becomes part of an automation when a trigger runs it with the necessary tools, inputs, and permissions.

Will a scheduled task run when my laptop is closed?

A cloud routine can. A local Desktop task depends on your machine and app being available. A CLI loop is session-based. Choose the runtime deliberately.

Is it free once I build it?

Not necessarily. Model usage, subscriptions, hosting, connected services, and maintenance can all have costs. Check the billing route before enabling recurring work.

Should I automate my whole second brain at once?

No. Start with one intake or review task. Preserve originals and test the output before allowing broad edits to your notes.

What is a good first automation?

A draft weekly brief is a useful candidate. It has clear sources, a repeatable format, and a result you can check before sharing.