Let AI do tasks for you

01What this module is about

So far the AI has answered, and you've done the doing. This module is about letting it do some of the doing too: looking things up across several websites, filling in a form, sorting files, or doing the same job every Monday.

AI that takes actions like this is called an agent. Agents save real time. They can also make real mistakes, because now a mistake is an action, not just words on a screen.

The answer isn't to avoid them. It's to start small, stay in charge, and check what they did. By the end of this module you'll have given an agent one small, safe task, and checked the result.

You don't need a paid plan. The task in this module works on any free plan, in a normal chat with web search. If your plan has an agent, you can use that instead.

02What an agent is

A chat gives you an answer. An agent does a task. It might:

• click through websites and gather information
• fill in forms
• read and change files on your computer
• run the same task on a schedule, like every Monday morning

These are the kinds of agent you'll find. Names and plans change often, so check what's on yours.

• Claude in Chrome works inside the Chrome browser. It needs a paid Claude plan.
• Claude Code works on files on your computer. It needs a paid Claude plan.
• ChatGPT Work can take actions for you. Switch from Chat to Work at the top of ChatGPT. It replaced ChatGPT's older "agent". On the free plan it's limited, and only in the desktop app. Don't count on it to stop and ask. In one test, reported by PCWorld, replying "continue" let emails go out without another check. Claude did the same in that test. Never say "continue" without reading what it's about to do.
• Task is Google's agent in Gemini. Click Task below the message box. You may have seen it called Gemini Spark. It needs a Google AI Pro or Ultra plan, and it isn't available in the UK, Europe, Switzerland or Nigeria.
• Scheduled tasks. ChatGPT, Gemini and Claude all have Scheduled in the sidebar. All three need a paid plan. Grok has Automations.

03Five rules for agents

1. Start small and safe. Pick a task where a mistake costs nothing. "Find and list" is a good first task. "Book and pay" is not.

2. Make it look and report before it acts. Ask it to tell you its plan first, and to wait for your yes. You'll spot a wrong turn before it happens.

3. Write "done" so you can check it. "A list of five walks with date, distance and start point" can be checked. "Find some good walks" can't.

4. You sign in, and you approve anything that sends, spends or deletes. Never give an agent your password. Sign in yourself if a site needs it. Some agents say they'll ask before a payment or a message, but Anthropic says plainly that you're responsible for what its browser agent does. Module 8 covers this.

5. Only schedule what works by hand. Run a task yourself, start to finish, a few times. Only when it's right every time should it run on its own.

04Text to copy

Use this brief for any agent task. Fill in the square brackets.

Task: [what you want done, in one sentence]
Where: [the website, or the folder on your computer]
Done means: [something you can check, like a list or a number]
Limits: Don't buy, send, delete or sign in to anything. Don't change any files.
First, look, and tell me your plan in three lines. Wait for my yes before you start.
When you finish, give me the result, and tell me what you did and anything you couldn't do.

Here it is filled in for the walking club:

Task: Find three circular walks under 6 miles that start within 10 miles of our village.
Where: Walking route websites that are free to read.
Done means: A list of three walks, each with its name, distance, start point and a link.
Limits: Don't buy, send, delete or sign in to anything. Don't change any files.
First, look, and tell me your plan in three lines. Wait for my yes before you start.
When you finish, give me the result, and tell me what you did and anything you couldn't do.

Save the result yourself, for example as walk-ideas.md in your project folder. A browser agent may not be able to save into your folder. Claude Code can. If you want it to save the file, change the Limits line to name that one file.

06Common problems

• It wanders off and does something else. Your task or limits were too loose. Make "Done means" more exact, and add what it mustn't do.
• It gets stuck on a sign-in page. Sign in yourself, then let it carry on. Never type your password into the chat.
• It says it finished, but it didn't. Check against "Done means", line by line. Ask: "Which parts didn't you finish?"
• It's slow. Agents often take longer than doing it yourself the first time. The saving comes when a task repeats.

07Keeping it current

• Keep the briefs that work. Save them in your project folder, so you can reuse them.
• Check scheduled tasks now and then. Websites change, and a task that worked last month can quietly start going wrong.
• Widen the limits slowly. Only let an agent do more once it has done less, well, several times.