The Open Hood.

Some people just seem to know how to run.

You've seen them. Same park, same Sunday, nothing special about the shoes, and yet they move differently. Lighter. Quieter. They come out of the bend at the same pace they went into it, and somehow they don't look like they're paying for it.

Talk to one of them and it's rarely raw talent. They know what's going on. Everyone knows the basics: move your legs, stay upright, breathe. The good runner knows why the back stays straight, and what a slump does to the breathing and the hips. They know why you land a certain way, what a heel strike sends up into your knees, and what a midfoot landing gives back to you as spring. They know where the energy comes from and where it leaks out.

You can copy what they look like. Straight back, short stride, light feet. You can buy their exact shoes. It won't turn you into them. Copying form is not the same as copying skill.

You don't have to become a sports scientist. You can just run, and you probably should work with a coach. But the people who are genuinely curious about how their body responds seem to keep finding another ten or twenty percent, long after everyone else has levelled off. It isn't that they train harder. They know which dial they're turning.

A house works the same way, only slower. Renovating one is dusty, over budget and mildly miserable, and it's also the one time the walls are open. You see where the pipes run, which wall carries the floor above, which wire goes where. So years later, when something leaks, you don't stand in the hallway reading a brochure. You know exactly who to call and what to tell them. Understanding the house is fun, though it's more than that. It stays useful for as long as you live there.

That brings me to the question under this whole piece. I keep asking it about AI, mostly of myself:

Do you just want to be an end user and a buyer of something? Or do you want to be a builder of something?

I'm asking now because the walls are open right now, and I don't think they'll stay that way.

What follows is my read. I've spent a lot of evenings with this, but I'm not an AI researcher, so take it as “this is what I've seen”, not settled science.

Why it's best at what its makers know

Most of today's AI was built by engineers and scientists. Making it is a scientific process and a creative one, and the people doing it have steered it toward what they decided to build, study and research. It's no surprise that it's best at the things closest to the people who made it.

There's a technical reason too, and I find it the more interesting one. A big part of how these models get better is reinforcement learning: the model tries something, the attempt gets graded, and the grade shapes the next attempt. For that to work, you need to be able to grade what the model did, cheaply and at enormous volume.

Code is gradeable. It runs or it doesn't. The test passes or it fails. You can check a million attempts without a single human in the loop. Now try grading a brand line. Ask three CMOs and you'll get four answers, and a fifth one the following Tuesday.

So the thing AI is really, really good at right now is coding. I don't think code is easier than marketing. I think code tells you when you're wrong.

It's now walking into your world

AI is now working hard at everything else: writing, marketing, business. It's not mature there yet. There are real use cases, and I use several of them every week. But the real gut of it, the stuff that makes you sit back in your chair, is still happening in the coding space.

That gap is the invitation. While AI is half through the door of your field, you can still see three things that will be much harder to see once it has fully arrived:

1. Why it responds the way it does. When it gives you a beige answer, that isn't random. It's often the average of every similar answer it has seen, because nothing in your brief told it to aim anywhere else. Once you notice when it falls back on the average, you've learned what it's actually doing.
2. What it gets right and what it gets wrong. Its mistakes follow a pattern, and the pattern is a map. Where does it flatten a nuance in your category with total confidence? Where is it surprisingly sharp? Today you can still watch that in the open. Once it's polished and packaged into a product, the seams get sanded away.
3. What scaffolding it needs. In coding with AI, the everyday words are scaffolding and harness: the instructions, tools, checks and memory you build around a model so it does reliable work instead of impressive one-offs. This is where it gets interesting for the rest of us. A harness is how you make an ungradeable field gradeable. Marketing doesn't grade itself, so you write the rubric. You define what good looks like, give the model the checks, and let it test itself against them before it hands anything back.

I think harnesses might become one of the most important things out there, and understanding the concept gives you an edge right now. I've written about it separately in The Agentic Nephew. The short version is that most of the difference between a toy and a colleague is the harness, not the model.

So here's the line I'd underline: the hood is open precisely because AI hasn't finished walking into your field yet. Once it has, you'll get the brochure like everyone else.

Why you can still catch up

At this point some readers start to feel late. Let me take that feeling apart, because I think it's mostly wrong.

In my day job, people join the team, people leave, people go on parental leave, and everyone takes breaks. In the evenings I build things myself, mostly vibe-coded, with the AI doing the heavy lifting. I'm not an engineer. Even so, I've watched people who joined the team late catch up, and quickly.

My own route started about two summers ago, in three stages, and each one took real time to learn, grasp and structure:

1. Prompting. I learned to write a good prompt, and back then a good prompt was a detailed one. It meant paragraphs of role, context, format and constraints, rewritten until the output stopped drifting.
2. Skills. I set up a harness and spent evening after evening writing skills: small, reusable written instructions the model could pick up when it needed them. (What that turned into, boiled down to a list, a manual and a clock, is in The Only-Me List.)
3. Manual orchestration. I learned to split a job into parts, sequence them, decide which instance did what, and give the whole thing structure by hand.

Today a frontier model with its slow, careful reasoning mode switched on will do a large part of that for you, as long as you don't fence it in too much. I call it lazy prompting: you write a high-level prompt that says what you want, without steering how it gets done. The model will take the time to spin up sub-agents to research the question. The coding harnesses now ship with a decent set of basic skills already wired in, so it doesn't have to reinvent what it does.

Two summers of my evenings, compressed into a default setting. I find that mildly humbling, and very good news for anyone starting now. The models have become the great equaliser.

So what do the concepts still buy you? They make you a somewhat better user. There's still a light user and a power user, and the difference shows up in four places:

• Pointing. A lazy prompt might work out by itself which skills it needs, try a few that don't fit, and get there in a couple of days. A power user points it at the right skills straight away.
• A shelf. A power user keeps a steady repository of skills that are tested, named and ready to use.
• A workshop. A power user runs an isolated workspace that gathers learnings across everything they do and turns repeated lessons into new skills.
• A spanner. A power user can update and change their own system without asking the frontier model to do it, which means fewer tokens and quicker work.

None of that is required. You don't need to know all these core concepts to be great with AI and have real impact. The light user with a good lazy prompt is already doing better work than most people were doing two summers ago with a page-long one.

And the bar keeps moving in your favour. My guess is that by the next frontier generation, three or six months out, even those power-user moves (isolation, learning across data, building skills, token efficiency) will mostly happen inside a conversation after a lazy prompt. The model will build its own shelf.

Two honest caveats

First, I could be wrong about how long the window stays open. The labs are working hard on grading the fuzzy stuff: rubrics, models that grade other models, human preferences collected at scale. If they crack it for writing and strategy, AI could walk into your field faster than I expect. To me that's a reason to start sooner, not to wait and see.

Second, catching up and understanding are not the same thing. The model will happily carry you through the door. It won't tell you why the door was where it was. That part you still have to want.

The two halves of this section sound like they contradict each other, but they don't. You can catch up, and understanding still matters. Shoes and a coach will get you round the park. Knowing why the back stays straight is what gets you the extra ten or twenty percent, and tells you when the knee is about to complain.