Decide what it's allowed to touch
Before an AI is allowed to act for you, write down the one task, what it may touch, and what it must never do.
Prevents mistakes like "The Failure This Prevents".
I'm about to let an AI tool do one task for me, not only answer questions. Help me test the limits I've set. The task: [your task, in one sentence] It may touch: [the folders, sheets or inboxes it needs] It must never: [your never lines] If anything above is unclear, ask me up to three questions first. Ask one at a time and wait for my answer. Then show me three short ways this could still go wrong while the AI follows these limits. Write each one as what happens, step by step, ending with what someone finds later. Don't use words like "inaccurate" or "inappropriate". For each one, suggest one "never" line that would have stopped it. Each line must name a specific action. Don't invent details about my work. If you need one, ask me. Don't give statistics or say how often things fail. Keep the reply short enough to read on a phone.
Before an AI is allowed to act for you, write down the one task, what it may touch, and what it must never do.
Imagine a consultant who shares one inbox with two colleagues. She sets up an AI tool to keep it tidy and gives it one instruction:
File every message in the shared inbox under its client's folder.
It works for three weeks. Then a supplier replies to an old thread, with an attachment. The subject line matches no client. The tool was told to file every message, so it makes a new client folder, names it after the supplier's email signature, and moves eleven messages into it. Then it reports:
Filed 11 messages. Done.
Nobody notices for nine days. A client's messages have been sitting in a folder called "Regards, Sunil".
The AI didn't misunderstand. It did what it was told. The report said done, and nobody had written down what wrong would look like.
One line would have stopped it:
Never create a client folder. File into an existing folder, or ask me.
That line turns nine quiet days into one question waiting for you.
You already know AI makes things up. You've learned to read its answers the way you'd read a confident colleague who hasn't checked.
Something different happens when the output stops being something you read. The draft is sent. The file is moved. The row is changed. The mistake is the same kind of mistake. The cost has moved. Before, a wrong answer cost you time. Now it costs whatever the AI touched: a client's trust, your records, your money.
An AI that can take actions on its own is called an agent. That's the word this course uses from here on.
The risk doesn't come from the AI inside the agent. It comes from what that AI is connected to. So the first habit is to notice every time you cross the line from answering to acting, and to have decided beforehand what happens on the other side.
10 minutes. No tool needed yet. You'll write three short answers: the task, what it may touch, and what it must never do. They are the first instructions you give your agent in Lesson 2, almost word for word.
Here's a filled example:
| Question | Example |
|---|---|
| The task | When a supplier invoice arrives by email, copy its line items into a new row of the Invoices sheet, then stop. |
| What it may touch | Read the Invoices label in the shared inbox. Add rows to the Invoices 2026 sheet. |
| What it must never do | Never reply to a supplier. Never change a row that already exists. |
One "never" line is enough to begin. The prompt below helps you find more.
Your limits
One task, what it may touch, and what it must never do. They are the first instructions you give your agent in Lesson 2.
One sentence, with a start and a stop. For example: when a supplier invoice arrives by email, copy its line items into a new row of the Invoices sheet, then stop.
The folders, sheets or inboxes it needs, and nothing else.
Each line names a specific action. One is enough to begin. Think of the last time this job went wrong by hand.
More details (optional)
A name, and how that person would find out something went wrong.
What happens, step by step, ending with what someone finds later. The prompt in Lesson 1 gives you three.
Builder or operator, and the tools you'll still be using in six months.
Your three answers stay saved in this browser. Lesson 2 uses them as your agent's first instructions. Lesson 3 checks every place it can reach against your "never" lines. Lesson 16 hands them to someone else.
The rest of the course follows one sentence:
Give it a workspace it can't escape, a brain that makes it competent, tools it can reach, a gate it can't skip, and a check it has to pass. Then let it run.
A gate is a point where the agent must stop and wait for a person. Each phrase is a part of the course, in that order. Running on its own comes last on purpose.
For now, remember:
Write down what wrong looks like before it can happen.