Claude Code Routines: What They Are, How They Work, Where They Win

Claude has been useful for a while. The catch was always the same. You had to start the conversation.

That changed on 14 April 2026. Anthropic released Claude Code Routines in research preview. A routine is a saved Claude Code configuration that runs on Anthropic's cloud, on a trigger, without you in the loop.

Close your laptop. The work keeps going.

The Shape of the Product

Every routine has four parts:

1. A prompt. Plain English. Self-contained. Specifies what to do, what success looks like, and what to do on failure.
2. Repositories. Cloned fresh from the default branch on every run. Claude writes changes to claude/-prefixed branches by default.
3. Environment. Network access is blocked by default. You allowlist what you need. API keys and tokens live here. The setup script is cached between runs.
4. Connectors. Your MCP connectors (Slack, Gmail, Linear, Notion, Calendar). All connected MCPs are included by default. Remove the ones you do not need.

That is the whole product.

Three Triggers

A single routine can attach to any combination of these:

Trigger	When It Fires
Schedule	Hourly, daily, weekly, or custom cron. 1-hour minimum interval.
API	A unique HTTPS endpoint per routine. POST with a bearer token. The body accepts a text field for dynamic context.
GitHub event	PR opened, closed, merged. Release published. 18+ event categories.

Stack them on one config. A PR review routine that also runs nightly to sweep stale PRs and can be fired from your deploy pipeline. One routine. Three entry points.

The Run Cap Is Tight

In the research preview:

Plan	Automated Runs Per Day
Pro (~$20/mo)	5
Max (~$100 to $200/mo)	15
Team / Enterprise	25

Manual Run Now clicks do not count. One-off natural-language /schedule runs do not count. Use them while you are debugging the prompt. Save your scheduled slots for the real work.

The Hack That Matters Most

The repo is not just for code.

Store your business context as markdown in a private GitHub repo and connect it alongside any working repos. Brand voice. Pricing. ICP definition. Customer personas. Standard operating procedures. Claude reads it all on every run without you re-explaining anything.

This is the single biggest difference between a routine that sounds like you and one that sounds like an AI. Most routines that fail are failing here. The prompt is fine. The business context is missing.

What Routines Are Not

Be specific. Routines replace cron-plus-glue-code where Claude was already doing the reasoning. They do not replace:

• n8n for multi-vendor pipelines moving data between Salesforce, HubSpot, Stripe, Slack.
• GitHub Actions for CI/CD pipelines.
• Anything needing sub-hourly cadence. 1-hour minimum is real.
• Local-environment work. Routines are cloud-only. They cannot touch tools on your laptop or services behind your firewall.

For everything in their lane, Routines win on reliability and zero infrastructure maintenance. For everything outside it, your existing layer stays.

Where to Start

Pick the smallest useful routine and ship it this week. A morning brief. A weekly SEO report. A daily inbox triage. Something that supports a decision rather than makes one. Something you can verify by eye in 60 seconds.

Then add the context repo. Then stack triggers.

The teams pulling ahead right now shipped a boring routine in May and a slightly better one in June. Complexity comes later, if at all.

---

This is the four-minute version. If you want the deep one, the field guide covers the full four-part architecture in interactive form, all three triggers in tab-by-tab detail, ten power-user hacks (including the SessionStart hook for getting local skills into cloud routines), six anonymised business use cases, and a copy-paste morning brief routine you can paste into Routines this week.

Read the Claude Code Routines Field Guide This is one piece of running AI as an operating system. The full picture is in our AI orchestration guide.

Once you have more than one routine firing, the next problem is where all that output lands. That is a surface, not a trigger, and it is covered in the AI command centre guide.

If you are working out how Routines fit into your operation, the IP Assessment is a ten-minute read on where the real return is.

Take the IP Assessment