Claude Code Routines: The No-Code Way to Automate Recurring Work

You keep meaning to send that Monday numbers email, or pull the week's support themes, or tag last week's expenses. It's a five-minute job that you skip because it's not urgent until it is. Claude Code Routines were built for exactly this kind of recurring, low-glory work — and you can set one up without writing a line of code.

Quick answer: Claude Code Routines are scheduled or triggered Claude Code sessions that run a plain-English task on a cadence — hourly, daily, or weekly. Anthropic launched them on April 14, 2026. You describe the job in normal language, pick when it runs, and Claude executes it in the background, so recurring work happens without you starting it each time.

What are Claude Code Routines?

A Routine is a Claude Code session that runs itself on a trigger you set instead of you opening a terminal and kicking it off. Anthropic shipped the feature on April 14, 2026, and the appeal is that the trigger can be a clock (every hour, every morning at 8, every Monday), an API call, or a GitHub event. You write the task the way you'd explain it to a competent assistant — "pull this week's Stripe charges, group them by category, and email me a summary" — and Claude runs that on the cadence you chose.

The reason this matters for a founder rather than an engineer: the task is described in plain English, not built as a flowchart of nodes. If you can write a clear instruction, you can write a Routine. That's the same "describe the job, don't build the machine" logic behind Claude Agent Skills for non-developers, applied to work that repeats on a schedule.

What can Routines run reliably, and what can't they?

Routines are strongest on fuzzy, judgment-flavored jobs that are annoying to build in a traditional automation tool. A Routine handles date reasoning ("last week," "the first Monday of the month"), light summarization, and pulling a few things together into a report far better than a rigid node-based flow does, because Claude reasons about the task instead of following a fixed path.

Where they're weaker: high-volume data plumbing and anything where you need to watch each step fail visually. If you're moving 50,000 records between systems every hour, a tool like Zapier, Make, or n8n is still the better fit — those win on throughput and on visual team debugging. The practitioner review at dominikgabor.com lands on the same split, and most small businesses end up running both: node tools for the heavy pipes, Routines for the judgment work. If you've hit the ceiling of your current stack, we cover that decision in when Zapier limitations mean it's time for an AI agent.

Reliability also has a hard cap you should know before you plan a schedule. As of April 2026, Anthropic limits daily Routine runs by plan.

If you're on Pro and want a Routine that checks something hourly, you'll blow through five runs by mid-morning. Match the cadence to the plan, or reserve your daily runs for the jobs that actually earn them.

The #1 day-one failure: OAuth token expiry

The most common reason a Routine works on Monday and silently dies the next Monday is OAuth token expiry, not a broken task. When you connect Google services (Gmail, Drive, Sheets, Calendar) to a Routine, the connection runs through Google Cloud — and if that project is in "Testing" mode, Google expires the token every 7 days. Your weekly Routine then fails without an obvious error, and you don't notice until the report you rely on stops arriving.

The fix is to move the Google Cloud project from Testing to Production mode before you trust a weekly schedule. dominikgabor.com flags this as the single biggest day-one failure, and it's worth doing the moment you set up any Google-connected Routine.

Do this once and the 7-day silent death stops. Skip it and every weekly Routine you build is on a countdown timer you can't see.

How to set up your first Claude Code Routine (no code)

Start with one boring, genuinely repetitive task — a weekly report is the cleanest first win. You'll describe the job in plain English, connect the tools it needs, set the cadence, and confirm the first run before you walk away.

Pick a task worth automating

Choose something you already do by hand every week and would notice if it stopped. A weekly support-theme summary, a Monday revenue snapshot, or a Friday "expenses I need to categorize" pull are all good candidates because the output is small, the value is obvious, and a mistake is cheap to catch. Avoid anything mission-critical for your first Routine — you want to learn the mechanics on a job where a bad run costs you five minutes, not a client.

Write the instruction like you'd brief an assistant

Write the task as a short, specific paragraph, naming the source, the transformation, and where the result goes. Vague instructions produce vague reports. Compare these two:

Weak:  Send me a weekly sales update.

Strong: Every Monday at 8am, read all Stripe charges from the previous
        Monday through Sunday. Group them by product. Give me total
        revenue, revenue per product, and the three largest single
        charges. Email the summary to me as plain text.

The strong version tells Claude the window, the grouping, the exact numbers you want, and the delivery format. This is the same discipline that makes any agent finish cleanly, which we go deep on in how to write a Claude agent prompt that finishes the job.

Connect only the tools the task needs

Give the Routine access to the specific services it uses and nothing else. If the task reads Stripe and sends email, connect those two — don't wire up fifteen connectors "just in case," because extra tool access inflates the context Claude has to reason through and makes runs slower and noisier. Connect three or four tools you actually use; that's the same restraint we recommend in the best MCP servers for founders guide.

Set the cadence and confirm the first run

Pick the trigger — daily at a set time, weekly on a chosen day, or hourly — keeping your plan's run cap in mind. Then run it once manually before trusting the schedule. Read the output. If the numbers are wrong or the format is off, fix the instruction and run it again. A Routine you've watched succeed once is a Routine you can forget about; a Routine you scheduled blind is a support ticket waiting to happen.

Common pitfalls when running Routines

Most Routine failures are boring and preventable, and they cluster around a few repeat offenders. Here are the ones that catch people in the first week:

Routines vs. the manual way vs. node tools

Doing this work by hand is fine until the day you're busy and skip it — which is every busy week. The point of a Routine isn't speed on any single run; it's that the job stops depending on you remembering.

For a founder running lean, the honest answer is usually a mix: a node tool for the heavy plumbing and Routines for the recurring work that needs a bit of thinking. Getting that first Routine running correctly is the hard part — the OAuth step and the "write it like a brief" instruction are where most people stall. Inside the free Claude Community there are copy-paste Routine templates for weekly reports, inbox triage, and expense pulls, so you can start from a working example instead of a blank instruction box.

FAQ

Do I need to know how to code to use Claude Code Routines?

You don't need to code to use Claude Code Routines. You describe the task in plain English — the same way you'd brief an assistant — pick a schedule, and connect the tools the job needs. The only technical step most people hit is moving a Google Cloud project to Production mode to keep the connection alive, and that's a handful of clicks in a settings menu.

How many Routines can I run per day?

Anthropic caps daily Routine runs by plan: 5 on Pro, 15 on Max, and 25 on Team or Enterprise as of April 2026. A single Routine that fires hourly counts as many runs, so an hourly task will exhaust a Pro plan's five runs before midday — match your cadence to your plan.

Why did my Claude Code Routine stop working after a week?

The most likely cause is OAuth token expiry. If your Google Cloud project is in Testing mode, Google expires the connection token every 7 days, and the Routine fails silently. Move the project to Production mode in the OAuth consent screen settings, then re-authorize the connection in Claude to fix it permanently.

Are Routines better than Zapier or n8n?

Neither wins outright — Routines and node tools solve different problems. Claude Code Routines handle fuzzy logic, date reasoning, and summarization better, while Zapier, Make, and n8n win on high-volume data movement and visual debugging. Most small businesses run both: node tools for the heavy pipes, Routines for the recurring judgment work.

What's a good first task to automate with a Routine?

A weekly report is the cleanest first Routine because the output is small, the value is obvious, and a bad run costs you nothing but five minutes to fix. A Monday revenue snapshot, a weekly support-theme summary, or a Friday expense-categorization pull all work well — pick something you already do by hand and would notice if it stopped arriving.