Everything can be automated with AI.
Why bother then?

Last week I tried to edit an Instagram reel with AI.

I am a software engineer. I have shipped products. I assumed this would take twenty minutes.

It took most of an evening.

Version 1 came back and it was fine. Not good. Fine. So I asked for faster cuts. Version 2 had faster cuts and worse text. Version 3 fixed the text and lost the rhythm. Somewhere around version 6 I realized I wasn't editing a video anymore. I was negotiating with one.

The AI wasn't the problem. It did exactly what I asked. The problem was that I didn't know what to ask.

We have been automating for 200 years

None of this is new. In 1801 the Jacquard loom automated weaving patterns with punch cards. In 1913 Ford's moving assembly line cut the time to build a Model T from about twelve hours to around ninety minutes.

But notice what Ford actually automated. Not "build a car." He automated a car that had already been designed, broken into steps, and specified down to the bolt. The line was fast because the system was clear.

Automation has always worked this way. It multiplies a system. It does not invent one.

AI feels different because you can type a wish in plain English and get something back. That makes it easy to forget the old rule still applies.

Give it a system and it multiplies the system. Give it a wish and it multiplies the variance.

Why it works when I write code

When I build software with AI, I rarely get the evening I had with that reel.

Why? Because I already know what the finished thing looks like. I know the constraints. I know the steps to get there. I know what's in scope and what isn't. When the AI produces something, I can tell in seconds whether it's right, and if it isn't, I can say exactly why.

That's a system. It lives in my head, built over years of doing the work badly and then less badly.

With video I had none of that. I knew I wanted it "edited." I wanted "some infographics" and "good animation." I wanted it to go viral.

But what is a viral video? How long is it? What happens in the first two seconds? Where do the cuts land? What does the viewer feel at second 8?

I didn't know. So I asked the AI to decide, and it decided differently every time. Each version was a guess. My feedback was a guess too, because I was reacting to what I saw rather than steering toward something I understood.

That's the loop. Vague intent, random output, vague feedback, repeat. Each round costs a context switch and a little more patience.

Variance is the symptom

The frustration of working with AI in an unfamiliar domain isn't a tooling problem. It's a knowledge problem wearing a tooling costume.

"Make me a viral video" is not a spec. It's a hope.

The variance machine

Generate six reels from the same prompt. Then tick a few rules, the way an editor would, and generate again. Watch what happens to the spread.

> make me a viral reel

hook in first 1s under 30s big captions, works on mute one idea per cut match a reference reel

Generate 6 reels

Usable on first try

0/6

Variance

Est. evening lost

3h

An expert editor would turn that hope into decisions before touching anything. A hook in the first second. Under 30 seconds. One idea per cut. Captions large enough to read on mute. A reference reel to match. Those decisions are the system, and once they exist, the AI can execute them almost perfectly on the first try.

The gap between version 1 and version 6 is exactly the size of the system you don't have.

It's the same for growing an audience

This goes well beyond editing. People ask AI to grow their Twitter or Instagram and get a pile of posts that sound like everyone else's.

Of course they sound the same. Growth over months isn't a prompt. It's a system: who you are writing for, what they already believe, what they need next, what you will post every week, and how you know if it worked. Without that, AI just helps you produce more noise, faster.

AI is leverage. Leverage on nothing is still nothing.

So why bother?

If everything can be automated, why learn anything?

Because the bottleneck moved. It used to be doing the work. Now it's knowing what the work should be.

Execution is getting close to free. Judgment isn't. The person who understands the domain well enough to write the system will get in one try what everyone else gets in ten. Over a year, that difference compounds into a different outcome entirely.

So the practical lesson is almost boring. Before you prompt, write the system down.

Do you have a system yet?

I can describe what "good" looks like, specifically enough that a stranger could judge it. I know the constraints: length, format, budget, platform, deadline. I can list the steps between blank page and finished thing. I know exactly who it's for and what they already believe. I have studied at least ten examples that worked and can say why.

0 of 5. You have a wish. Expect version 6.

If you can't answer those, that's your real task, and AI can help you learn it faster than ever. Study ten reels that worked. Ask why. Turn the answers into rules.

Then automate.

Ford didn't start with the assembly line. He started with the car.