Introduction

Have you ever calculated how much time you waste on repetitive tasks that feel mindless?

You wake up, open a few websites, copy and paste data, and create a report to send to your team. Every Monday, you gather numbers from various backends to fill in a static Excel sheet for your weekly report. You find a great article and want to save it, but it ends up never being opened again.

You can’t afford an assistant and don’t want to learn programming to set up a server. These “mental chores” grind you down like sand in your shoes, draining your energy.

Until I discovered a severely underrated feature hidden in Claude—Claude Routines. This translates to automated routines.

What can it do? In short: You write the tasks for it, and it executes them daily on time, without you needing to turn on your computer, rent a server, or even think about it. Once completed, it neatly stores the results in your specified location or sends them directly to your WeChat/DingTalk/email.

Today, I’ll break down a complete guide on building your personal automation army from scratch. Whether you’re a complete novice who doesn’t know what code looks like or a developer wanting to understand its limits, you’ll be able to get started immediately.

What Is It and How Is It Different?

Previously, if you wanted AI to help automate tasks, you had limited options: either set up a server and write scripts (exhausting), use expensive low-code platforms (costly), or rely on rigid “bots” that only follow fixed scripts (inefficient).

Claude Routines breaks down all these barriers.

It runs on Anthropic’s own cloud and is included in your existing subscription. The key is that it doesn’t come with a rigid “script brain” but rather full Claude intelligence.

What does this mean? It thinks for itself while working. If it finds missing data, it will find a way to fill in the gaps; if one method fails, it will try another; if it feels something is off, it will check back. This is a truly intelligent agent that can think, not just a robot that executes commands.

How Does It Think and Work?

Before you get started, you need to understand its operational logic. This will determine how effectively you can use it.

Every time it runs, it follows this process:

Trigger (time to run/signal received) → Spin up a new cloud computer → Clone your GitHub repository (this is its hard drive) → Read your instruction manual → Work (call APIs, process data, write files) → Store the results back to GitHub → Destroy the cloud computer, leaving no trace.

The key point: Every run is a fresh start. The computer is new, and the environment is new.

This means two huge benefits: first, you never have to maintain a server, and you don’t have to worry about the environment crashing unexpectedly. Second, each run is clean, and the junk from the previous run will never affect the next one.

However, this also means that if you want it to “remember” what it did last time, you must have it write the memory to your GitHub repository. For example, you can have it save today’s data as yesterday_data.json, and when it runs tomorrow, it will read this file to have “memory.”

Three Wake-Up Methods: Ensuring It Shows Up When You Need It

Method 1: Scheduled Wake-Up (Most Common)

Tell it in plain language: “Every Monday at 9 AM” or “Every day at midnight.” It supports standard Cron expressions, allowing advanced users to control it precisely. Note that the minimum interval is one hour; it cannot run every five minutes.

Use Cases: Daily briefs, weekly reports, morning news summaries, scheduled data scraping.

Method 2: API Wake-Up (Most Powerful)

Each routine you create will automatically provide you with a unique URL (API endpoint) and a key (Token). Any external system capable of making a web request can wake it up.

Imagine this: a new order comes in from your online store (Shopify sends a signal) → wake up the routine → the routine tells Claude to draft a personalized thank-you email. Someone fills out your form (Typeform sends a signal) → wake up the routine → the routine tells Claude to analyze the requirements and generate a preliminary plan.

Use Cases: Event-driven automation, connecting Claude with any existing software that supports Webhooks.

Method 3: GitHub Event Wake-Up (For Programmers)

Events like PR submissions, new Issues, or code pushes can trigger it. Automatic code reviews for PRs, automatic categorization replies for Issues, and automatic documentation updates are all possible with this method.

Step-by-Step Setup: From Zero to Running

Step 1: Write the “Instruction Manual” (The Most Critical Step)

Remember a core principle: Claude cannot ask you questions while running automatically. Your instructions must be complete, clear, and unambiguous, just like writing a job description for a genius employee.

Don’t write: “Help me see what’s happening in the crypto world today and summarize it.”

Instead, write: “You are a crypto market analyst. Your tasks are: 1. Call the CoinGecko API to fetch data for the top 10 DeFi tokens by market cap; 2. Compare today’s data with yesterday’s and calculate the percentage change; 3. Generate a Markdown report that includes a summary, data table, and anomaly analysis; 4. Save the report in the /outputs/ folder; 5. Update /memory/yesterday_data.json. The API key should be read from the environment variable CG_API_KEY.”

Advanced Tip: Starter Prompts

In the UI, you can simply write: “Run the /skills/daily_brief.md skill file in the repository.” All complex details can be packed into that file. Clean, maintainable, and version-controlled.

Step 2: Connect to GitHub Repository (Its Hard Drive)

You must provide it with a GitHub repository. This is where it stores memory, outputs, and reads instructions. It’s recommended to use a dedicated, streamlined repository; don’t throw your entire company codebase in there, as it will slow down performance and waste resources.

Step 3: Store Passwords in the “Cloud Vault” (Environment Variables)

API keys, email passwords, DingTalk robot Tokens… never write these in the instruction manual (that would be like broadcasting your passwords online). Go to the routine settings and store the keys in the “cloud environment variables.” It will read them automatically during execution.

Step 4: Set Network Permissions

By default, it can only access websites certified by Anthropic. If you want it to scrape a niche website or call a wild API, remember to change the network permission from “trusted” to “all” or “custom.”

Step 5: Trim Down Unused “Plugins” (Connectors)

The routine will automatically include all plugins you’ve previously connected (Gmail, Slack, Notion, etc.). Make sure to delete all those that won’t be used for this task. A routine for writing daily reports shouldn’t have access to your email. Removing unnecessary plugins reduces the risk of accidental operations.

Step 6: Run Once Before Scheduling

Click “Run Now” at the top and monitor the real-time logs. Check if it found the instruction file, read the passwords, and completed all steps. Most failures are due to incorrect file paths, mismatched password names, or network blocks.

The Secret to Keeping Its Intelligence Online: Skill Architecture

Don’t pile complex instructions in the UI. Create a /skills/ folder in your GitHub repository, placing Markdown files inside. Each file represents a “standard operating procedure.”

The file structure should follow this template:

• Purpose: Clearly state what this skill does in one sentence.
• Preconditions: What passwords and plugins are needed.
• Input: Where the data comes from.
• Process: What to do first, second… the more detailed, the better.
• Output: Where the file is stored, in what format, and whether to send messages.
• Exception Handling: If any step fails, should it log or send an alert?

In your repository’s root directory, place a CLAUDE.md file that outlines global rules: tone style, date format, brand persona. Every time it runs, it will read this file to ensure all outputs carry your essence.

The Real Power: Building Memory for Continuous Improvement

Most people use automation to just take snapshots: run once today, produce today’s output; run once tomorrow, produce tomorrow’s output. These two are entirely unrelated.

The real power lies in ensuring that today’s run builds on the results from yesterday’s run.

Mode 1: Difference Tracking

At the end of each run, write the current state to /memory/snapshot.json. On the next run, read this file to calculate the changes. Your daily report will no longer be a dry list of numbers but rather “What happened in the last 24 hours.”

Mode 2: Accumulative Logs

Each run appends new data to /memory/history_log.json. After a week, you’ll have complete trend data. After a month, the AI can directly help you analyze long-term patterns.

Mode 3: Sliding Window

Only keep data from the last 7 days. Each run uses these 7 days as context (for example, to avoid repeating topics when writing content outlines), then adds today’s new data while discarding the oldest entry.

What If It Messes Up?

First, in every skill file, enforce the addition of an “Exception Handling” section: “If any step fails, log it to /outputs/errors/ and send a message to the DingTalk group #AI Alert.” It’s better to have it proactively alert you than to wait for you to discover the issue.

Second, if a failure occurs, troubleshoot using this tree:

• Did it complete the run? If not, check the logs to see where it got stuck.
• Is it a password/permission issue? Look for messages like “403,” “file not found,” or “credentials.”
• Is it a network issue? Look for messages like “domain blocked” or “timeout.”
• Is the instruction too vague? The output will be nonsensical.
• Did it exceed the quota? It may stop halfway through.

Security Rules: Don’t Let It Run Wild

An intelligent agent that can autonomously access your APIs, send emails, and write files needs strict security measures.

• Principle of Least Privilege: Only give API keys the minimum permissions necessary (read-only is often sufficient).
• Minimize Explosion Radius: Only attach necessary plugins, send messages only to a dedicated test channel, have it draft emails without sending them, and manually review for a week before fully enabling.
• Set Repository to Private: There’s no reason to make your automation routine repository public.
• Add a Final Lock for Irreversible Operations: Write in the skill file: “If output quality is below standard, only write to file, do not send. And alert for manual intervention.”

Not Sure Where to Start? Let Claude Help You Design

If you finish reading and still feel confused: “I know this is powerful, but what exactly should I have it do?”—Let Claude help you identify needs.

Open a regular chat window (not the routine interface) and copy this text:

“I want to use Claude routines to automate my work and life. I am [your role/profession]. A typical week for me includes [describe what you do, what annoys you, and what is valuable]. The tools I commonly use are [list software]. What I want to optimize is [saving time? More accurate information? Or ensuring nothing is missed?]. Please help me list 10 automation ideas ranked by time savings, implementation difficulty, and reliability, explaining how to do each.”

It will provide you with a highly customized list. Pick the one that excites you the most, and have it generate a complete skill file and setup checklist. You just need to copy and paste.

Conclusion

Most people use Claude as a remote control—press a button, and it moves.

Routines allow you to turn it into an autopilot system. Set the destination, and it plans the route, responds to conditions, arrives on time, and then messages you: “Boss, I’ve arrived.”

With 25 free runs daily (5 for Pro, 15 for Max, and API triggers don’t count against the quota), you can outsource all those annoying “mental chores”.

Start with the smallest task. For example, every morning at 7 AM, have it send you summaries of titles from a few specified websites to your WeChat. After a week of stable operation, set up the next task.

Stop doing things manually that machines can do better.

What’s the first task you want to automate? Let’s discuss in the comments; you might find existing skill templates!