Routines: scheduled cloud agents

Set up a scheduled Claude Code routine that reviews your website repository and leaves a report for you to inspect. Start with a narrow reporting job, run it once, and enable repetition only when its output is useful.

Before you start

• A Pro, Max, Team, or Enterprise claude.ai account with Routines available. Routines are in research preview; organizational Owners can disable them.
• A connected GitHub repository that the routine can clone, plus a cloud environment suited to the task.
• A schedule in your local timezone and a small repeatable job, such as checking navigation links in repository files. Routines consume subscription usage.

Want to do it yourself?

Difficulty

8 out of 10

Estimated completion time

1–2 hours, plus a scheduled trial run

Skills you’ll need: Configure a cloud task, permissions, and schedule, then inspect a delivered run and practise pausing it.

I’ll do it myselfI’d rather get help

How this helps your brand

A recurring navigation review can surface website changes that need attention. A checked first report and narrow scope make the routine easier to maintain without assuming repository analysis proves the live site works.

1. 1

   Step 1: Write a self-contained job

   Choose something you can judge from a report, such as finding internal navigation targets that have no matching page in the repository. Say what to inspect, what counts as evidence, and what to do when the answer is uncertain.

   Use the prompt below for a first run. It asks for a report, not edits or external messages. Avoid “maintain my website” as a first routine: you would have no clear way to tell whether it finished the intended work.

2. 2

   Step 2: Create the routine on the web

   Visit claude.ai/code/routines and select New routine. Give it a descriptive name, paste your job into the instructions, select a model, and add the repository it needs.

   Choose the cloud environment. The run starts from a fresh clone, so unsaved files on your laptop are not inputs. If the review depends on a file, make sure it is present in the repository the routine will read.

3. 3

   Step 3: Set a schedule and review access

   Under Select a trigger, choose Schedule and a suitable cadence, such as weekly. Check the local time you entered. Scheduled starts can be late, so choose a review that does not depend on an exact minute.

   Review Connectors before saving. Connected services are included by default and can allow writes without a permission stop during the run. Remove all connectors for this repository-only example, and give the environment only the network access the job needs.

   Illustrated exampleA schedule with a small, clear scope

   ↻

   Example routineWeekly navigation review

   Schedule

   Monday · 9:00 a.m.Your local timezone

   Repository

   northline/websiteFresh clone for each run

   Connectors

   NoneRepository-only review

   1. Scheduled start9:00 a.m. local
   2. Cloud reviewMay start later
   3. Read the reportIn the run’s session

   Choose a review that can wait. A scheduled time is not an exact-minute guarantee.

4. 4

   Step 4: Run once and read the actual report

   Select Create, open the routine’s detail page, and choose Run now. Open the resulting run as a session and read what Claude did.

   Compare its findings with at least one known route in your repository. Check that the report names files and clearly marks patterns it cannot resolve. A green run status indicates no infrastructure error; it does not prove the requested review succeeded.

   Illustrated exampleRead the evidence behind a completed run

   Illustrative session reportInternal navigation review

   Run completed

   Scope: repository files in app/ and components/. No live browser test.

   1 Confirmed missing

   components/footer.tsx:18 links to /pricing. No matching page or redirect was found in this example repository.

   2 Unresolved pattern

   components/projects.tsx:32 builds /projects/${slug}. Available slugs depend on external content; skipped.

   3 Known route

   /contact has a matching app/contact/page.tsx. Open that file yourself to check this finding.

   ✓

   Check the evidence before repeating. A completed run tells you the run finished. File references and an honest skipped-checks list make its report useful.

5. 5

   Step 5: Fix the cause of an incomplete run

   If Claude cannot clone the repository, reconnect GitHub and confirm the selected repository is accessible. If it reports a blocked network request, inspect the requested host and decide whether that access is actually needed.

   For a necessary domain, edit the routine’s environment and add the domain under Custom network access, then save. Use the menu beside the routine name and Edit to correct its prompt or configuration. Run it again and check the report before trusting the schedule.

6. 6

   Step 6: Keep or pause the schedule

   Leave the routine enabled only after the manual run provides a report you can use. Note when you expect the next report and check that scheduled run too.

   Use the on/off switch at the top of the routine detail page to pause it when the task is no longer needed. Check your account usage rather than assuming unattended runs are free.

Review internal navigation in [repository] for [pages or folder in scope]. Read repository files only. Identify links whose target appears missing and cite the source file and target for each. Separate confirmed missing routes from dynamic patterns you cannot resolve. Do not claim a live browser test was performed. Do not edit files, commit, push, open pull requests, send messages, or publish anything. Return a short report in this run’s session with findings, evidence, skipped checks, and one recommended next action. If required access is missing, report the blocker and stop instead of guessing.