I’ve been using these AI models for a couple of years now, and Claude Code — the one that can actually reach into my files and run things — hard for about six months. For most of that time I thought I was using it to get work done. I was. But that wasn’t the main thing that was happening, and I only see it now, looking back.
What I was actually doing was learning. Not how to code, exactly — though a little of that rubbed off. I was learning to see my own work as information that has a shape.
The idea that cracked it open was structured versus unstructured: the difference between a fact that lives in a defined field — a date, a vote, a dollar amount, a category — and a fact that lives loose inside a paragraph. That sounds like a database distinction, and it is, but in a media field it turns out to matter enormously. In journalism we live in the unstructured, and we fall in love with it — it’s where the nuance and the interpretation and the mystery are. That love is the craft, and I wouldn’t trade it. But a lot of the information people actually need from us isn’t mysterious at all. It’s straightforward: when, how much, who voted which way. And our instinct is to dress even that up as prose. We take a three-hour meeting where a budget got set and four votes were cast and a tax rate was proposed — all of it structured, every bit of it a fact with a shape — and we flatten it into one blob. Once it’s a blob it can’t update itself, it can’t connect to the last six meetings, and a reader can’t follow the single thread they actually care about. The structure was always in there. We just threw away the part that made it usable.
Learning to pull the structured facts back out — the vote into a vote, the date into a date, the candidate into a record — is what made everything else possible. The structured parts are the parts a machine can hold and a page can fill itself from. The unstructured part — what it meant, why it matters, what to make of it — is the part that stays mine. It’s the old line between rules and judgment, but you can only sort the two once you can see which is which. The way I picture it now: let the automation pour the foundation and raise the frame — the load-bearing parts that have to be straight and the same every time — so the work left to me is the architecture and the style and the ornament. What the house is for, how it feels to stand inside it. The part only a person can decide.
And I learned to see it by building it, badly. The real on-ramp for me was n8n — one of those visual automation tools where you drag boxes onto a canvas and draw lines between them: this happens, then this, and here’s the data getting handed from one step to the next. Looping information, learning where code would do what I needed instead of an LLM node, debugging things that broke — wiring up those flows was the first time I could see a procedure instead of just performing it — a chain of discrete steps with information moving through, each step expecting a particular shape. Eventually I had a fleet of pretty good automations that pulled and edited text, but looking at them I started to sense the raw code underneath and wanted to understand it — what else could I do?
Handing tasks to Claude Code and watching exactly where it succeeded and where it fell on its face was how I learned to tell the rules from the judgment. So a couple of years of conversations that felt like getting things done were really a couple of years of fieldwork on my own operation. A lot of it was throwaway — a one-off task, a prompt I never used again. But the throwaway parts were the experiment. They were how I found out what could be systematized and what couldn’t.
The interesting part is what’s happening now. I’m going back through all of those pieces — the experiments, the half-built things, the prompts that actually worked — and nailing the final ones together as real code. And here’s what I didn’t expect when I started: the best of them don’t use the AI at all anymore. Just code. Clean, consolidated, routine — a thing that posts every story to social, a thing that pulls the lodging-tax numbers each month, a thing that fixes the same garbled name in a transcript the same way, every single week. Because the LLM, it turns out, was scaffolding. You raise it around the thing you’re building and work from it while the structure takes shape — and when the structure can stand on its own, the scaffolding comes down. The AI helped me find the recipe; once the recipe stopped changing, I nailed it together in code and pulled the scaffolding away.
That’s the opposite of dependency, which is the thing everyone seems to be worried about. And I’ll be honest: taking the scaffolding down is the part I enjoy most. LLMs are amazing, but they’re still a little unpredictable — and I want my building to last. A piece of code does the same thing today that it’ll do next year; an AI might surprise you either way. So the work I’m proudest of is the work that ends with no AI left in it. Scaffolding is supposed to come down — that isn’t the tool failing to be needed, it’s the tool having done its job. When an automation graduates from “the AI does this” to “this just runs,” the scaffolding has come down on something that can finally hold its own weight.
People like to argue about whether AI will replace what someone like me does. Wrong axis. The framing I keep coming back to is Thomas Ptacek’s: even if you think LLMs lower the ceiling in some domain, if you’re being honest you have to admit they raise the floor. It was never going to raise the ceiling — a good reporter still covers a meeting better than any machine, and I’ll keep saying that to anyone trying to sell you otherwise. What it did was raise the floor: it let one person build the boring, permanent structure a one-person newspaper could never afford to build before. And for a small paper, the floor is where you live or die.
So I spent six months thinking I was getting help with my work, and it turns out I was getting an education in my work — learning to see it clearly enough to take the parts that don’t need me and make them run on their own. The scaffolding was never meant to stay up. The building is what I keep.
Leave a Reply