How to get really good at using AI

Metacognition built under mismatch becomes a strength in systems that reward clear, explicit thinking.

1. Spend a lifetime examining your own thinking and behavior until you accidentally become really good at metacognition.

2. Make it your profession to understand other people’s thoughts, patterns, and weird little edge cases.

3. Open an AI chat, type “I need to think this through,” and just be yourself.

It’s a joke. Sort of.

But if you opened an AI chat for the first time and were somehow immediately good at it, without a course or framework, or without really understanding why, this might be about you.

Metacognition means thinking about your thinking. People keep telling me I’m good at it and I have no idea what to do with that. It’s like being told I’m good at hearing. I watch my own thinking happen constantly, whether I want to or not, and I can’t turn it off or imagine what it would be like without it (which honestly reflects poorly on my UX design skills).

I turned that into a career designing things for other people’s brains while still figuring out my own. And when I look at what actually happens when I use AI well, it’s never a technique I picked up from somewhere. It’s knowing what I want and being able to say it. Or not knowing yet, and being willing to stay in that uncertainty long enough to figure it out through conversation. It’s being able to notice when something went wrong and tracing back to why. Holding a model of my own thinking and the AI’s interpretation at the same time, and feeling when they don’t line up.

That’s metacognition. And it’s not a new idea that it matters for AI. A human-computer interaction paper from 2024 made the same argument. People who are better at it hand over less to AI and are pickier about what comes back. That part is established.

What I’m really curious about is who’s already been thinking this way, and what that tells us about what minds actually need from the tools we build.

What all this watching-my-own-brain has taught me to do is take something that makes sense inside my head and reformulate it so it makes sense to someone or something that doesn’t think like me. I do it in conversations with others, when I design interactions, and when I write. It’s the constant quiet work of translating between my inner experience and whatever I’m trying to communicate with.

That’s literally what prompting is.

Taking a vague idea and making it linear enough for something else to work with. Nobody calls it that when you do it with humans, but every person who’s ever had to question their own experience just to figure out what they’re actually asking for has been doing prompt engineering their whole life. They just didn’t always have an LLM on the other end.

I have what you might call a peer diagnosis. Every time I write about how my brain works, a certain group of people show up. Complex thinkers, with or without an ADHD and/or autism diagnosis, who are excellent at building systems and seeing how things connect while being almost uncomfortably self-aware. They recognize something in what I describe. And I recognize them back. However, the public health care system here decided I function too well to qualify for an assessment. And I’ve been sitting with that for a long time, but it wasn’t until I started writing this that I realized why it really bothers me so much.

The functioning is the metacognition. The constant monitoring, the translating, the adjusting, the building of elaborate internal systems so that from the outside it looks like you’re fine, even though you’re not. You’re running a second operating system alongside everything else and hoping nobody notices the lag. But a clinician looks at the output, she works, she parents, she communicates, and concludes there’s nothing to assess. It’s a funny little feedback loop happening here.

The better your metacognition works, the less likely you are to qualify for help with the thing that made you build it in the first place. The more you notice the gap between your inner experience and the world, the better you get at translating to fit in.

And I believe that exact skill is what makes me good at AI.

Systems get built around one kind of brain, they become invisible to the people they fit, and then nobody questions them because they just work. For some people. When the machinery doesn’t fit your brain, you have to take it apart. Why does this system work for everyone else and not for me? What is this person actually asking for beneath the words they used? Why can’t I just follow instructions and adapt? It’s exhausting and it comes with a cost. Masking, burnout, years of manually operating a machine that everyone else seems to drive on instinct. A lot of neurodivergent people have been doing this translation work forever, out of necessity.

But what all that taking-apart builds, over years, is an explicit model of how thinking and communication actually work. Not the implicit kind that hums in the background, but one you constructed by hand because the default version didn’t work for you.

And then along comes AI. A language model that rewards explicit instruction and constant self-correction. The coping mechanism I built because nothing worked right finally meets a system that treats it as a feature. A lifetime of compensation and the machine says: more of that, please.

When the output comes back wrong, you don’t get frustrated, blame the model and retype the same thing louder. You read it and feel where the mismatch is. That’s not what I meant, but I can see why it got there. You adjust and try again from a different angle. Not because someone taught you a technique, but because that’s what you do with every human system that doesn’t understand you on the first try. For better or worse, you assume you are the problem because that’s what you are used to.

AI might be the first mainstream tool where the default is no longer the advantage.

The research I can find on neurodivergence and AI is about accommodation. Using it to cope and fit in, despite your wiring, in order to do what neurotypical brains already do. Finally a tool that can help you fight your natural way of being in order to keep up with the unreasonable standards of modern life (sarcasm, for some reason I’m fluent in it, that might be the thing I need to understand next.)

Even Jurgen Appelo’s well-researched argument that AI is an interface adapter is still asking how to make the existing system more bearable. And this is where I keep getting stuck. Because even when the argument is kind, it still treats AI as a prosthetic that helps you survive a world that stays the same.

I’ve lost count of how many times I’ve referenced this post by now, but Jessie Mannisto wrote about people whose thinking naturally mirrors how LLMs process. She writes about people who think in constellations, and how that shape can look a lot like the way transformers move through language. LLMs don’t just tolerate that kind of mind. They meet it.

Jessie is talking about wiring. I’m building on that, and looking at what happens when the wiring doesn’t match the demands of the environment it’s in. A lifetime of translating yourself and rebuilding your own thinking into something other people can follow turns out to be the exact skill the machine rewards.

I don’t want to make this into a superpower narrative, though. The metacognition I’m describing was expensive. It was built in years of constant self-monitoring and not fitting in. Metacognition can be learned. So maybe neurodivergent, high-masking brains just get there earlier.

But I don’t think the motivation is evenly distributed. The drive to translate comes from somewhere specific. It comes from not being understood. If the world has always made sense to you, if systems worked the way your brain expected them to, if people generally got what you meant on the first try, you never built this muscle because you didn’t need to.

So when the chat works, it feels like relief. And also like grief. Not because the model is magic. Because I should not have had to become this good at translation just to be understood.

It’s just thinking about thinking. Out loud. And translating myself yet again.

— Emma

Also, the song I had on repeat while writing this:

Subscribe for more overthinking about AI, design, neurodivergence, and what else we could build.