The best AI course I've taken is the thing I built for myself

I haven’t learned AI by studying AI. I’ve learned it by giving myself things to make.

A website. A research system. A GitHub repository. Editorial frameworks. Experiments. Articles. Small tools. Workflows that connect some of those things together.

Collectively I’ve called it Editorial Intelligence.

I originally thought of it mainly as an idea about content: how editorial judgement, evidence and narrative might become more valuable as AI makes producing content easier.

It’s been doing something else as well.

It’s become my AI training ground — not in the conventional sense. There’s no curriculum and I’m not collecting certificates.

I’m learning because I keep running into actual problems.

The questions only turn up when you’re building

How do I give an AI enough persistent context to be useful?

How do I stop it producing generic writing?

What belongs in the repository and what doesn’t?

What should be automated, and what shouldn’t be?

Where does human judgement actually matter, as opposed to where I’d like to believe it matters?

How do research, retrieval and synthesis change when a machine can work across all three at once?

How do you build something with AI without gradually handing over the thinking that made it worth building?

None of those questions arrived from reading about AI. Every one of them arrived from trying to make something work and discovering it didn’t.

That last question in particular has no useful answer in the abstract. You only find the line by getting close enough to it to notice you’ve crossed it.

I’ve written separately about why an experimental edge needs to sit outside the day job — small, cheap, reversible, nobody’s transformation programme riding on it. That still holds and I won’t repeat it here.

What I hadn’t appreciated is what the same setup does for two other problems.

”AI skills” is becoming a phrase that means nothing

It’s on every job advert and half the LinkedIn profiles I read. It can mean anything from having a subscription to having built production systems.

I’d rather be able to point at something.

This is what I built. Here’s how it works. Here’s what I learned. Here’s the bit I changed when it didn’t work, and here’s the thing I published and then withdrew because the argument turned out to be wrong.

That last category matters more than it looks. A portfolio of things that worked proves you can pick winners or tell a good story afterwards. A record that includes the failures proves you were actually running the experiment.

Because much of this is public and version-controlled, the record isn’t reconstructed. It’s just there, with timestamps, including the parts I’d have quietly tidied up if it were a CV.

A career hedge that doesn’t require quitting anything

I don’t know exactly what my profession looks like in five years.

Nobody does, and I’ve become suspicious of people who sound certain about it.

Trying to predict it precisely seems less useful than increasing my ability to adapt to whatever turns up. So this has become a hedge — but an unusually cheap one.

It doesn’t require me to leave my job. It doesn’t require me to become an AI consultant. It doesn’t even require Editorial Intelligence to become a business.

If it does, good. But the project has already returned something regardless: somewhere to practise the future before I have to work in it.

What this doesn’t tell you

Two things I should be honest about, because this is exactly the kind of piece that turns into advice it hasn’t earned.

It works because I already had somewhere to stand. Fifteen years of editorial judgement gave me a domain to build a system around, and problems worth solving. Somebody starting out doesn’t have that, and “build your own thing” is much harder advice to follow when you don’t yet know what the thing should be about. That’s the same gap I’ve written about in the ladder problem, showing up from the other direction.

You’re reading about the project that survived. Plenty of things I started went nowhere and were never written about, because nothing draws you to write up an abandoned experiment. Every account of learning-by-building has this problem, including this one.

So take it as a description rather than a prescription.

The part I think actually matters

I keep coming back to it because I enjoy it.

That sounds trivial next to arguments about capability and career strategy, and I don’t think it is.

The best learning system I’ve found isn’t the one with the best curriculum, the clearest progression or the most credible certificate at the end.

It’s the one that makes me want to open the laptop again tomorrow.