← Learn
Playbook · 10 steps · checked 4 October 2026 · 0 adopted

Routines

At most ten steps, in order of how much they help. Each step is from Anthropic's own pages, linked. Re-checked against those pages every week; anything that can't be confirmed is marked unverified rather than removed.

  1. 01

    Write the prompt as a complete brief, with what success looks like

    A routine is a saved prompt, repositories and connectors that runs on its own, with no permission prompts. Anthropic says the prompt is the most important part: it must be self-contained and explicit about what to do and what done means. Pick the model in the prompt box; every run uses it. (routines)

  2. 02

    Pick the right kind of schedule for the job

    Routines run in the cloud from a fresh clone, so your machine can be off, but the minimum interval is one hour. Desktop scheduled tasks run on your machine with your local files, but only while the app is open and the computer is awake. /loop (for example /loop 5m check the deploy) polls inside an open session, and recurring loops expire after seven days. (scheduling options, Desktop tasks)

  3. 03

    Create and manage routines from the terminal with /schedule

    Describe it in plain words, such as /schedule daily PR review at 9am, or a one-off like /schedule in 2 weeks, open a cleanup PR that removes the feature flag. Claude asks follow-up questions, then saves it to your account. /schedule list, /schedule update and /schedule run manage existing routines. It needs a claude.ai subscription login, not an API key, and does not work inside a cloud session. (routines)

  4. 04

    Click Run now, then read the transcript

    A green status only means the session started and ended without an infrastructure error, not that the task worked. Blocked network requests, missing connector tools and task failures show up in the run itself. From the CLI you can ask, for example, /schedule why did my nightly review do nothing this morning? (Claude Code v2.1.227 or later). (routines)

  5. 05

    Remove every connector and permission the routine does not need

    All your connectors are included by default, and Claude can use every tool in them, including writes, without asking. Anything the routine does appears as you: commits carry your GitHub user and messages use your linked accounts. The Default environment only reaches an allowlist of common domains; other hosts fail with 403. On Pro and Max, keep API keys as API credentials rather than environment variables, which anyone using the environment can see. (routines)

  6. 06

    Schedule a few minutes past the hour

    A run set exactly on the hour, such as 9:00, can start several minutes late; 9:07 starts closer to time. Times are entered in your local zone. For a custom interval, pick the nearest preset, then set a cron expression with /schedule update; anything more often than hourly is rejected. (routines, changelog)

  7. 07

    Guard your branches on GitHub, not in the prompt

    Claude pushes to claude/-prefixed branches unless the prompt says otherwise. Use branch protection rules or rulesets to limit where runs can push; a rule your connected GitHub access can bypass will not block a run. If your GitHub connection lapses, runs are skipped for up to 72 hours, then the routine turns off until you reconnect and switch it back on. (routines)

  8. 08

    Wire alerts and CI to an API trigger, and tell the prompt to use the payload

    Add the API trigger on the web, then generate the token and store it straight away; it is shown once. The optional text field (up to 65,536 characters) arrives labelled as untrusted data, so the saved prompt must say to act on it, for example "Investigate the alert described in the routine-fire-payload block". There is no idempotency key, so a retrying caller starts extra sessions. (routines, fire API)

  9. 09

    Filter GitHub triggers so only the right pull requests start a run

    Install the Claude GitHub App on the repository first; /web-setup grants cloning access but does not enable webhooks. Triggers cover pull request and release events, and filters such as "is draft is false" or a label skip runs you do not want. matches regex tests the whole value, so use .*hotfix.* or the contains operator. (routines)

  10. 10

    Watch your usage and pause routines you no longer need

    Runs use your subscription the same way interactive sessions do; check claude.ai/settings/usage. There are also hourly caps: 100 scheduled runs per account, and 30 Run now, API fires and one-off re-runs per routine. Use the on/off switch to pause a routine without losing its set-up. (routines, fire API)

Unverified

What Anthropic's own pages disagree on, or could not be confirmed.

  • The launch blog post gives daily limits (Pro up to 5 routines a day, Max 15, Team and Enterprise 25). The current routines docs list only hourly limits plus subscription usage. Treat the daily figures as possibly out of date. (blog)
  • The routines page's curl example sends an anthropic-beta: experimental-cc-routine-2026-04-01 header and says the endpoint ships under it; the API reference says the header is optional. Sending it appears harmless either way. (routines, fire API)
  • Routines are in research preview, so behaviour, limits and the API may change without much notice.