This post was written with Claude, and that's not a disclaimer.
I used it to pressure-test my argument, check for fallacies, research anything I needed to verify, and run the final text against a style guide I built, with Claude's help, from my own previously published writing. What you're reading is my thinking, structured by me, expressed in my voice. Claude helped me get there, but didn't get there for me.
The default voice problem
People tried AI for writing, hated the results, and wrote off the tool. Usually with some version of "it doesn't sound like me." That's accurate, because out of the box, it doesn't. It sounds like everyone else who used the same tool without giving it any direction.
The people who can now spot AI-written content immediately are often the same ones who gave up on it: they got burned by the default voice, stopped using it, and now recognize the pattern everywhere. They're identifying a specific default, not making a judgment on the technology. A hammer doesn't drive screws either, and that's not the hammer's fault.
What the smell actually is
The tells are consistent enough that they become hard to miss once you've seen them.
There's the opening gambit: a sweeping scene-setter that says nothing. "In today's world, content is more important than ever." There's the false balance move, where every observation gets immediately hedged into mush: "While X is true, it's also important to consider Y." And there's the em-dash infestation, so common it's been called the ChatGPT dash, where every sentence that could breathe gets interrupted, subdivided, and interrupted again.
The problem is with the default, not the technology itself. The model was trained on a lot of content farm output and SEO-optimized blog posts, or at least that's what the defaults suggest, and without a countervailing voice, it produces more of the same. Avoiding AI doesn't solve that; giving it something to work against does.
The style guide is the thing
I built mine from my own published writing. Claude analyzed it, I pushed back on the analysis, we refined it together. Now when I use Claude to help with a post, it knows what I actually sound like: short paragraphs, no em-dashes, no filler phrases, no motivational closer asking you to leave a comment below. It knows what words I don't use and what structural patterns I reach for, so it has something to work with other than the average of the internet.
The result is writing that sounds like me, because the constraints come from me. I bring it my argument and ask it to help me shape it, spot the holes, and clean up the language, rather than feeding it a topic and asking for a post.
The thinking and the judgment are mine. Claude is handling what Claude is good at, which is not having opinions about the thing I'm writing about.
The guide itself isn't finished either. Every post refines it a little, and catching what slipped through is part of how it gets better, which is how any tool improves.
I ran this post through five AI detectors, and four were highly confident no AI was involved; the fifth hedged. The goal was to produce something genuinely mine in terms of tone and argument, not to fool a detector, and apparently that's what the style guide delivers.
The same logic applies to developers
I wrote earlier that a programmer's value is in the thinking, the proximity to the problem, the judgment built up over years of being wrong, not in the typing. AI tools didn't change what the job was; they changed how much time the job now gets, because the typing is handled.
Writing isn't different. If you have something to say and you know how to say it, you now have a tool that handles the parts that were never the point, and that's time back.
The actual question
The only question worth asking is whether the thinking was yours. A lazy, derivative post written entirely by hand is still lazy and derivative, and a sharp, honest post written with significant AI involvement is still sharp and honest. The tool doesn't determine the quality; the process does.
What I've described here is one process. It works for me because I spent time making it work for me, which is, it turns out, how tools work.