How to Make AI a Habit for One Task

The people who use AI every day aren't more disciplined than you. They set one recurring task up once and never start from a blank prompt again. This week: how to pick the task, build the standing setup, and run it for two weeks straight.

Laptop with orange project sidebar beside clipped reference pages and a tabbed memo

Every survey on AI at work says the same thing right now: a lot of people have tried it, and a much smaller group actually works with it every day. I don't think the daily crowd is smarter or more disciplined than the rest of us. They just stopped starting from a blank prompt. They picked one thing they make every week, set it up once, and now the setup does the remembering. That's the whole habit, and you can build it in an afternoon.

// The Lesson

Set one task up once, then run it every time.

❝

Pick one deliverable you produce every week, build a saved project with standing instructions, reference material and one finished example, then run it every time for two weeks and fix the setup, not the draft, when it misses.

Gallup's latest read on this, a May survey of 22,573 U.S. employees, puts it plainly: 52% now use AI at work, but only 15% use it daily. Most of the distance between those two numbers is the blank-prompt problem. When every session starts from nothing, you have to re-explain who you are, what the document is for and what a good one looks like, and by the third time you'd rather just write the thing yourself. The daily users don't re-explain. They open a setup that already knows. Here's how to build one.

Step one: pick one deliverable you already owe every week

Not a category. One specific thing with a name and a due date: the Monday status memo, the brief before the weekly customer call, the summary of last week's support tickets, the first draft of the sales follow-up. Two tests. It has to come up at least once a week, because a habit needs reps. And it has to have a shape you could describe to a new hire: roughly this long, for this reader, in this order. If you can't describe the shape, pick something else. Don't chase the most valuable task you do. Chase the one that comes around most often, because that's the one the setup will earn back fastest.

Step two: build the standing setup once

Every major assistant has a place for this now. Claude calls them projects, ChatGPT calls them projects too, and Gemini calls them Gems. Same idea underneath: a saved space with its own instructions and files, so every chat inside it starts with the context already loaded. Put three things in it. First, standing instructions, written the way you'd brief that new hire: who reads this, how long it runs, what to lead with, what to leave out. Second, the reference material the task always draws from: the template, last quarter's numbers, the account list, the style notes. Third, one finished example you were actually happy with. This is the one people skip, and it does more work than the instructions, because it shows tone and structure instead of describing them. Anthropic's help center and Google's Gems page both describe the setup the same way: instructions plus files, carried into every chat in the space. Budget an hour. It's a one-time hour.

Step three: run it every single time for two weeks

Now the rule that makes it a habit. For the next two weeks, every time that task comes up, it starts in the project. No exceptions, including the Tuesday you're in a hurry and it'd be quicker to just type it. The first run will be a bit off. That's expected. What you do next is the whole lesson: don't fix the draft and move on. Fix the setup. If it ran long, the instruction gets a length. If it led with the wrong thing, the instruction says what comes first. If it made up a number, the reference file gets the real one. Every miss becomes one line in the instructions or one file in the folder, and by the second week the draft comes back close enough that you're reading instead of rewriting.

Why doesn't everyone do this already? Two honest reasons. Building the setup feels like a detour when the memo is due at four, and the blank prompt genuinely is faster the first time. It's the fifth time that gets you, when you're pasting the same background in for the fifth week running. The other reason is that people quit after a rough second run and decide the tool can't do it, when what actually happened is the setup was missing one line. So here's the move: pick the task and build the project before Friday, then put a two-week reminder on the calendar. Skip it and next Tuesday's memo starts from a blank page again. Do it and it starts from a draft that already knows your reader, your numbers and your last good one.