---

What an automation means here

Forget flowcharts and nodes you drag around with a mouse. What happens here is simpler: you describe a recurring task in plain language, decide when it should run, and say where the result goes.

The only difference from a normal chat is repetition. You can reassemble a competitor report by hand every week. Or you write down once what it should look like and just read it on Mondays.

What you need

Four things, nothing else:

• Claude does the actual work
• Cowork is the desktop side, where tasks run on a schedule and reach your files
• MCP servers are the connections to your other apps, things like calendar, storage or a database
• Claude Projects is where your templates and style rules live

No stack of subscriptions, no chain of connectors, no code.

Step 1: Create a project as home base

Set up one project per area, not one per automation. In my case that is a handful: content, clients, research, filing.

The project instructions hold everything Claude needs to know about you so that you never type it into a prompt again: how you write, which formats you use, how your files are named, what you never want to see.

Your first step: create one project and write three sentences about your writing style into it. That is enough.

Step 2: Describe the task in four parts

This is the core. All 45 automations consist of the same four parts. Leave one out and the result becomes useless.

1. Trigger. When should this run? Every morning at seven, every Friday, or whenever you drop something into a folder.
2. Source. What should Claude work from? Your inbox, a folder, a spreadsheet, a website.
3. Task. What exactly should happen? Not "summarise this", but "pull out the decisions, plus open questions and who does what by when".
4. Destination. Where does the result land? A file, a folder, a draft in your mail client.

Here is the industry digest from the list:

Trigger:     daily at 6am
Source:      the twelve sites on my reading list
Task:        filter for relevance to AI in marketing,
             condense to five bullets,
             nothing older than 24 hours
Destination: new file in my research folder, date in the filename

That is the whole automation. No code, no configuration.

Step 3: Run it manually once

The step almost everyone skips. Before you put anything on a schedule, run the task by hand once and look hard at what comes back.

Most of the time the first run shows the task was too vague. You get a polite summary instead of the decisions you wanted. So you sharpen part 3 and run it again. Two or three rounds is normal.

Only when the result looks good enough to forward unread is the automation finished.

Step 4: Put it on a schedule

Now you attach the trigger. Daily, weekly, monthly, depending on how often the task actually comes up.

Two rules I have settled on:

• Less often beats too often. A competitor report that arrives daily gets ignored by day four. Weekly, you read it.
• Schedule it before your working day starts. The result should be waiting when you sit down, not popping up in the middle of your focus.

Step 5: Add the next one

From here it gets fast. The second automation costs you ten minutes, because the pattern is set and the project already knows your style.

Start with the task that annoyed you most this week. Not the one that sounds most interesting technically.

Common mistakes

• The task is too vague. "Write me a summary" returns padding. Say which four things it has to contain.
• No fixed destination. If the result stays in the chat, it is not an automation, it is a conversation you forget.
• Everything at once. Set up twenty automations on the first weekend and by Monday you have twenty outputs nobody checks. Start with three.
• Never looking again. Sources change, folders get renamed. Ten minutes once a month covers it.
• Letting it send unchecked. Anything that goes outward, so emails, proposals, replies, stays a draft. The last click is yours.

Next steps

Pick three automations from the list that match your actual week. Build them with the four parts from step 2. Let them run for a week and throw out whatever you do not read.

What survives is your real list. It ends up shorter than 45 and actually in use.