How Far Do You Let It Run?

In the early days of AI, people talked about prompt engineering as if getting the magical words in the right order would make the AI do the right thing. For a split second, that was probably true.

It isn't true now. What actually makes you good at this is how far you let the AI run without you checking every word. That distance has a name: trust.

Maybe you're going through an AI draft line by line tonight. It's probably fine, but you read every word anyway, because it goes out under your name. That's the smart move, not a sign you're behind. You just haven't handed enough over yet to find out the machine can carry it.

Picture learning to drive with an instructor in the passenger seat. At the start they've got a hand near the wheel and a foot over the second brake, reading every move you make. Months later they're looking out the window, relaxed. They didn't stop paying attention. They learned to see trouble coming, so they could let you go further.

That distance answers a pile of worries that only sound separate: which model is best, what the perfect prompt is, whether this takes your job. Under all of them sits one question: how do you let it run further without getting burned?

A book called Prediction Machines made the point that stuck with me. When AI makes guessing cheap, your judgment gets worth more. The machine guesses what comes next; you decide whether it's worth doing. That call stays yours.

So you never flip from checking to not checking. Further along, the machine shows you what matters instead: a flag, a number that's drifted, the thing that surfaces before you'd have thought to ask. You hand over more, never on the dodgy boyfriend saying "trust me" while you look away, but because you can see what it's doing.

How the distance grows

What follows is one line with six marks on it. It isn't a set of levels you climb, and there's no mark you're behind on. It just shows where you are and what's a little further out.

1. Basic tasks. You drop a one-off into a chat and read every word before it goes anywhere. "Write a reply to this email." "Summarise this report." Asking a 20-word question instead of a 3-word one counts too. You're checking everything, which is right at the start. Even so, the boring first draft is off your plate, the first slice of your week handed back.

2. Context. You save a few files so you stop re-explaining yourself: your brand voice, your product range, who your customer is, how you like to work. Now the answers come back sounding like you, and you let it go a little further, because you've stopped checking that it remembers the basics.

3. Skills. You take the clever thing you worked out once and move it out of the prompt into a small script that does the exact same job every run, instead of a fresh roll of the dice each morning. To work with your numbers, you write the code that does the sums; you don't hand the model a thousand rows and hope.

Anthropic, the people who make Claude, had a prompt 400 lines long trying to explain every step in words. They moved the actual logic into a skill, and the prompt dropped to about 15 lines. Because it then ran the same every time, they stopped needing to re-check it.

4. Systems (the brain). Skills, context, and your data, wired together across the business. This is what I call the brain. Now it drafts, sends, and updates without you sitting in the middle. If you run an agency, this is where a junior's first-draft client report can go to the brain instead of to you, and come back close enough to finish. It's an engine, not a magic button, so I give it a simple test before it acts on its own:

• How bad is it if this is wrong?
• How sure am I it's right?
• Can I undo it?

A draft saved for review passes all three, so the brain handles it alone. A refund sent to a customer fails the last one, so that stays mine.

5. Runs without you. The brain stops waiting to be told. It works in the background and brings things to you before you ask: the renewal that's due, the client account that's drifting before the client emails you about it, the reply that's ready to send. You've handed over not just the doing, but the deciding-what-to-do. That stays safe on two conditions: it knows what it may do alone and what it must ask about, and it writes down what it did, so any morning you can open that list and see what ran overnight. When everything's fine you don't hear from it; it only reaches for you when something's off or a costly call needs a yes.

6. Self-improving. The system gets better on its own, week to week: noticing what worked, sharpening its skills, without you rebuilding it each time. Maybe this mark worries you, so let me be plain: this is not the machine doing everything. You handed over the job of getting better at the job, never the job of deciding what good looks like, because that's the thing it tests every change against. Get that definition wrong and it gets confidently worse, fast. So your hand stays on the wheel, right out at the far mark.

How you move further

Once you can see roughly where you are, the next move is smaller than it looks. You don't leap to the far end. You hand over one more thing, and only once the last handover has earned it.

So find your spot on that line and take the next step on purpose:

• If you're still dropping one-off tasks into a chat, save your first context file: who your customer is, how you write, how you work. Watch the answers start to fit you.
• If the answers already fit, take the thing you redo by hand every week and turn it into a skill, so it runs the same way every time.
• If you've built a few skills, wire them to your real data so they work across the whole job, not one task at a time. That's the brain.
• If the brain already drafts and updates for you, give it the rules and the logbook it needs to run overnight.
• If it already runs without you, start checking its own work against what good looks like, so it can sharpen itself.

You're not racing anyone, and there's no prize for getting to the end first. You just keep asking what you can hand over next, and the machine keeps handing your time back.