Tell AI “build this” and code comes out. Run it, and it more or less works. Convenient. But at some point it started feeling off. The code runs fine — I just don’t know why it looks the way it does.
Back when I typed every line myself, I naturally knew. Now AI rewrites five files at once, and I scroll through the diff going “yeah, looks fine,” and move on. The app keeps growing, but the map in my head doesn’t grow with it. The cheaper implementation gets, the harder it actually is to learn the technical side.
Skip the review and you quietly hand over control
I first put this into words while brainstorming YouTube topics. I was toying with something like “Building an app solo with AI — the real reason it was harder than expected,” and started writing down what that difficulty actually was.
- AI writes the code, but if you can’t set direction, it wanders off course
- If you don’t know what you want, AI can’t help you either
- In the end, requirements, verification, and prioritization matter more than the code itself
So I landed on: the solo developer in the AI era is less a coder and more a small-scale PM and reviewer. The weight shifts from “being good at building” to “being good at directing and verifying.”
I’d written all that down neatly, but in actual practice, verification was the first thing I’d skip. If the screen looks roughly right, I’d call it done and move on. And things I let slide like that come back to bite you later.
That’s exactly what happened building myWiki. Early on, I just dumped document bodies into a database column. It worked fine, and there was no obvious reason to revisit it. Then I went to add a “text upload” feature and hit a wall — the whole point of the product is that AI reads and greps my writing folder-by-folder, and that’s impossible if the body is locked inside a column. I’d missed a mismatch I should have caught way earlier. In the end I tore the whole storage layer down and rebuilt it file-based. If I’d actually looked hard at that column decision even once, back when I made it, none of this would have happened.
Even when something looks finished, actually run it yourself. AI is what makes you skip that obvious step — because everything it produces looks so smooth. The moment you skip the review, control quietly slides over to AI. It’s your app, and you’re the one who understands it least.
Blogging forced that review
Knowing review matters doesn’t help if you’re building solo — there’s no built-in mechanism forcing it. There’s no real reason to interrogate yourself with “wait, why did I build it this way?” If it runs, you move on.
Starting to blog changed that.
Explaining something in writing means going back and looking at “why did I build it this way.” Why I skipped the database, why this folder structure, why this pipeline order. Building the reasoning back up for a reader produces two things. One is confidence — oh right, that’s why I did it this way. The other is a hole — mid-explanation you catch yourself going “wait, why did I do this, this seems off,” and a problem you’d missed surfaces.
It’s the same principle as setting the angle during blog-brief. Give AI just a topic, and you get the generic version anyone would write — something about Notion, something about Obsidian. That’s the article anyone could have written. But hold onto “why did I actually make this call” while you write, and the act of writing re-verifies your own decision. The differentiation in the writing and the review of the code come from the same act.
So the real payoff of blogging was never the published post. It’s the process of reviewing in order to write it. The content is a byproduct. The actual thing is “time spent understanding again what I built."
"Doesn’t that take forever, though?”
This is where most people stop. Development’s already eating all my time — where’s the time for blogging too?
Fair point — but it changes once you set up a pipeline. That’s what this whole series has been about. I built a skill pipeline that turns a one-line description into a draft, and set it up so AI manages the entire blog with nothing but text files, no database. So getting one post out the door doesn’t actually cost much hands-on time.
The review still has to be me. AI pulls a draft out of raw conversation logs, and I read it and catch “hm, that’s not quite what I think.” That check is the review. The labor of drafting drops out, and what’s left is only the judgment that actually needs me. That’s precisely what the pipeline does — strip out what doesn’t need me, and leave only the review that does.
So — I recommend it
If you’re building something with AI, I’d recommend explaining what you built in writing at least once. Not for the traffic. You don’t even need to publish it.
The moment you look back to explain it, the decisions you’d let slide land back in your hands. You come back to a state where you actually understand why the code looks the way it does. Blogging is the thing that forces that, every time — and without it, I’d have handed my own app off to something else a long time ago.
The content isn’t the point. The review habit is the point. The post is just the mark that habit leaves behind.