How to set up AI that runs itself.
The shift from prompting AI to handing it a goal and walking away. Here's the setup I use, step by step.
Liked this? New guides land here first. No spam, ever.
Prompting keeps you in the loop: type, wait, fix, repeat. The fix is to set AI up so you give it one goal and it runs itself. Five things make that work: tell it the outcome not the steps, save how you like things once, connect it to your tools, let it run on a schedule, and have it check its own work. Here's how I set each one up, with a prompt you can copy for every step.
Give it the goal, not the steps.
For months I prompted AI like everyone does. I'd type a request, it wasn't quite right, I'd fix it, type again. I was the bottleneck in my own setup. The thing that changed it was small. I stopped describing the steps and started describing the result I wanted.
"Write my Monday email" became "Here's my week. Draft the emails that need to go out, and flag the ones that need a decision from me." When AI knows the destination, it can work out the route. Your job moves from operator to editor, and that's a much nicer job.
Here's the result I want: [describe the finished thing, not the steps to get there]. Ask me up to 5 questions if you need to, then do it. Tell me anything you had to assume.
Tell it once how you like things.
AI forgets you between chats unless you tell it not to. So I wrote it a short note about me. How I like things done, my tone, the stuff I always want and the stuff I never want. I saved it once. In Claude that's the memory, or the instructions inside a Project.
Now I don't re-explain myself every morning. It's the difference between a brand new temp every day and someone who already knows the job. You write it once. It pays you back every time after that.
Remember this about how I work and use it from now on:
What I do: [your role or business]
My tone: [3-4 specific words]
I always want: [formats, length, the way you like things]
I never want: [the stuff that's off for you]
From now on, default to this unless I say otherwise.
Plug it into your actual tools.
AI that can only talk is half the value. AI that can reach your inbox, your calendar and your files can actually do the thing. In Claude, connectors let it read your Gmail, see your calendar, pull from your drive. Once it's connected, "what's on my Monday and what can wait" stops being a question you answer and becomes one it answers for you.
One honest note. Only connect what you're comfortable with, and have a look at the privacy settings first. You're in charge of what it can see.
Look at my calendar for next week and my unread emails. Tell me what actually needs me, which ones you can draft replies for, and what I can safely ignore.
Let it run on its own.
This is the part that feels like magic. Instead of starting AI every time, you set it to start itself, on a schedule or a trigger. Mine drafts my content for the week every Friday morning, before I'm even up. I open my laptop to a finished first draft. I'm not doing the task. I'm reviewing the result.
The deepest version of this lives inside Claude Code, and the apps are adding more of it all the time. You don't need the deep version to start. One recurring job, set to run on its own, is enough to feel the difference.
Every [day and time], do this for me: [the recurring job, described as an outcome]. Draft it, save it, and have it ready for me to review when I sit down.
Make it check its own work.
Left alone, AI will hand you its first draft and hope for the best. So I make it grade itself before it reaches me. I ask it to do the work, then check that work against what I actually asked for, and fix anything that's off. It catches its own mistakes, so I'm not the one finding them at 11pm.
This one step is what makes walking away safe. It's the difference between an assistant you have to watch and one you can trust with the whole job.
Before you give me the final version, check your own work against my original goal. List anything that's missing or off, fix it, then give me the clean version and one line on what you changed.
Don't try to automate everything at once. Pick the one task you do every week that you resent the most, and set up just that. Get it running, learn to trust it, then add the next one. A system you actually use beats a perfect one you never finish.
One last thing.
All of this has a techy name. People call it loop engineering, and the power-user version lives inside Claude Code. But you don't need any of that to start. You need one goal, one note about how you like things, and the willingness to let AI finish a job without you hovering over it.
That's the whole shift. I'm still building mine out, bit by bit, around a full-time job and three kids. You can start yours today, with one task.