I Keep Deleting the Same AI Sentence. I Think I Finally Get Why It's There.

Last week I opened a draft one of my AI agents had written and there it was again: "It's not about being everywhere. It's about being unmissable." Not the first time. Not the tenth. I have genuinely lost count of how many times I've deleted that exact sentence shape out of AI-written work before it goes anywhere near a client.

I built a small AI marketing team to run alongside me — separate roles for content, SEO, paid media, and reporting, all wired into the same system Pip manages. They're good at their jobs. They are also, on a sentence level, wildly predictable. Ask any of them for a paragraph and I could tell you before opening the file that it'll come back with a heading, at least one bulleted list, and somewhere in there, an "it's not X, it's Y."

I have an actual checklist for this

I'm not exaggerating the checklist part. I keep a real list — twenty-five items, plus a few I've added since — of phrases and patterns that read as AI-written the second a human eye hits them. The X/Y contrast. Exactly five bullets per section. "The catch?" as a transition. "Navigating the complexities of." I run every draft against it before anything goes near a client's inbox, because owning a checklist and actually catching every instance turned out to be two very different skills.

For a long time I filed the whole pattern under "annoying habit to edit out" and left it there. Then I started asking a different question. Instead of treating the pattern as decoration, what if it's closer to wiring: a byproduct of how the model organizes the information, not proof it's phoning it in.

Bullets aren't a style choice. They're a signal.

Watch what happens when you feed a model one messy sentence instead of a structured one. Something like: conversions are down, but CTR looks fine, so it's probably the landing page or the tracking, and we should check both before touching the ads. Ask a model to work with that, and it hands you back a problem line, a not-the-problem line, a short list of causes, and a call on what to check first — four distinct pieces where there used to be one long thought running together. Every relationship between those ideas just got made explicit. That's not laziness. That's the model doing exactly what training rewarded: turning one tangled idea into pieces it (and anyone downstream) can actually use.

The "it's not X, it's Y" problem, specifically

The phrase that shows up more than any other on my list is the X/Y contrast. "It's not about working harder, it's about working smarter." "This isn't a technology problem, it's a systems problem." It reads like a LinkedIn post mid-breakdown, and I understand exactly why my checklist flags it on sight. But look at what the sentence is doing underneath the cringe: it names the wrong interpretation, then the right one, in a single beat. For a model resolving what a sentence means, that's an unusually clean signal. We've just collectively decided, after reading a thousand versions of it, that the signal reads as slop.

Maybe good prompting was never about magic words

A lot of what got branded "prompt engineering" over the last couple of years — the special phrasing, the exact spot to put your instructions, the little incantations — might just be the slow, collective discovery of which structures make information easiest for a model to actually use. I didn't get better results out of my agents by finding the right words. I got better results the week I stopped writing them paragraphs and started writing them what are basically spec sheets: the role, the scope, the exact output format, a couple of examples of what good looks like, and a named list of what not to do. That's not a prompting trick. That's the same document I'd hand a new hire on day one. I'd just never bothered writing it down until a piece of software finally made me.

The loop none of us are fully steering

Here's the part I can't quite let go of. These models learned to write by reading us. Then they started writing at a scale no single person ever could. Now a meaningful share of what's on the internet either passed through a model or was shaped by someone who reads a lot of model output, whether they clock it or not. Which means the sentence I keep deleting from my agents' drafts this month is quietly auditioning to become normal English next year — not because it's good writing, but because enough of us will have read it enough times for it to stop registering as strange.

The bullet point might be less a formatting choice and more the closest thing a model has to showing its work.

So does the AI actually want bullets?

No. Not in the way my son wants a cookie before dinner, a wanting I recognize instantly because I feel it myself most days around 3pm. A model doesn't have a preference sitting in there somewhere, quietly rooting for the numbered list. What it has is a strong, repeated pull toward certain structures, because those structures kept working across an almost unimaginable pile of training text. That's a real pattern worth paying attention to. It's just a different kind of thing than a preference, and I think that difference matters more than it sounds like it should.

What I actually changed

Here's what's different in how I write things now, prompt or otherwise. I name the relationship between ideas instead of implying it — problem, not-the-problem, cause, priority, stated plainly instead of buried in one long sentence. I write down what not to do as specifically as what to do, because a model, like most people I've managed, can act on "skip the executive summary" a lot faster than it can infer it from tone. And I stopped treating structure as the thing you bolt on at the end to make a draft look finished. I start with it now, because it turns out the shape of the information was the actual work the whole time.

I still don't know if any of this makes me better at talking to AI or slightly worse at talking to humans. I'll report back the next time I catch myself sending my sister a bulleted list by text.