Generic output is a missing input.
Ask a model to write about your topic with nothing else in the prompt and it will do exactly that. It writes about your topic, for a reader it has never met, in the voice of everyone who has already covered it. The result comes back correct, on time, and completely interchangeable. That combination is what average looks like when it shows up on schedule.
The fix isn’t a better model, and it isn’t a cleverer prompt. It’s a layer, and it’s the first of the five that make up a working setup: a short description of how you sound, kept in a file the tool reads before it writes, plus a checklist that catches whatever slips past it anyway.
Why the default converges
A model produces the most probable continuation of what came before. Handed a topic and no other constraint, the most probable continuation is the middle of everything written about that topic. Not a bug. A description of the job you gave it.
The compounding version is worse than the single draft. Each generic paragraph arrives looking acceptable, so you accept it. It’s fine. Fine ships. And fine accumulates into an archive where a reader can’t tell one of your pieces from a competitor’s, which is a positioning problem that started as a prose problem nobody could see at the sentence level.
There’s a second-order effect worth naming here. The more you work with a tool that produces the average, the more your own sense of a baseline slides toward the average, because the drafts you read most often are the ones it wrote. Taste calibrates by exposure. Feed it enough competent middle and competent middle starts reading like your standard.
That drift is invisible per draft and obvious across six months. Read three of your own pieces at six-month intervals and you’ll know immediately which end of the range you’re standing on. I’ve written about where to keep AI out of the work entirely. This is the other half: what to do about the work you keep it in.
What a voice layer actually contains
Four things. None of them takes long to write, and the fourth does most of the work.
Who you write for. Not a demographic. The reader’s actual situation, expertise, and what they’re currently stuck on. “Solo founders past the getting-started phase, fighting inconsistency, short on time, sceptical” is usable. “Entrepreneurs aged 25 to 45” is not.
What you’re positioned against. The thing you’re the alternative to. This is the part most people skip, and it’s the one that produces edge. If your file can’t name what you refuse to sound like, every draft drifts toward the thing you refuse to sound like.
How you sound. Sentence rhythm, paragraph ceilings, contractions, whether you use “I” or “we”, how a piece opens and closes. Specific enough to check. “Conversational but professional” is a mood. “Full contractions, sentences under thirty words, never open with a rhetorical question” is a rule.
What you never publish. A list of banned words, phrases, and constructions. Hype vocabulary, the transitions that give an AI draft away, the tics you’ve caught yourself producing. This is the shortest file and the highest-yield one, because it’s the only part of the layer that fails loudly. Everything else is a preference. This is a tripwire.
I’ve written before about building the knowledge base a tool writes from. The voice layer is the smallest version of that, and the one to build first.
The gate is the part nobody builds
Files aren’t enough. Instructions are context, not enforcement, and a model that reads your voice file will still hand you something flat on a long enough timeline. So the layer needs a second half: a checklist that runs between the draft and everything downstream.
Mine is one page. It checks the hard fails first, the ones that mean rewrite rather than edit. Banned construction present. Wrong opening technique. Any claim I can’t source. Anything that reads like a coach. Then the softer blocks: rhythm, transitions, whether the closing does one of the things closings are allowed to do here.
The order matters as much as the contents. Hard fails first, since there’s no point assessing the rhythm of a paragraph that’s about to be deleted. Then the opening, which is where a piece either sounds like a person or sounds like a template. Then rhythm, transitions, the closing. Format last, because format is the cheapest thing to fix and the least likely to be what’s actually wrong.
A fail costs a rewrite of the offending passage, not the whole piece. That distinction is what keeps the gate from turning into something I quietly stop running. If clearing it meant starting over, I’d abandon it inside a month, and a gate you’ve stopped running is worse than one you never built, because you’ll still believe you’re holding a standard you’ve actually dropped.
The rule that makes it work is that it’s a gate, not a score. A draft doesn’t advance because it scored well somewhere else. The generic tools all optimise for a number, and a number is exactly the thing that lets a competent, flat, on-brand-adjacent piece through. A draft that scores 78 in my voice beats one that scores 92 and reads like it was assembled. Clearing the gate doesn’t mean shipping, either. That’s a separate rule.
What the gate caught last week
A draft came back structurally right. Correct length, right sections, sourced, on-brief, nothing factually off. It also had almost no contractions in it. Every sentence was “it is” and “that is” and “does not.”
No single line was wrong. The whole thing read like a manual.
The gate caught it because the voice file says contractions are full and spoken. Without that line written down somewhere a tool reads, I’d have shipped it, because at the sentence level there was nothing to object to. The tell only exists at the article level, which is the level you stop being able to see once you’ve read the same draft four times.
The second catch that week was one clause: a sentence that explained something by ruling out what it wasn’t, in a construction sitting on the banned list. Small. Also the exact kind of small that repeats itself two hundred times across an archive until it becomes the thing your writing sounds like.
Why it has to be a file
A habit forgets. Not usually, and not on a good day. It forgets on the Thursday where the piece is late, the energy is gone, and the draft in front of you is good enough that arguing with it feels like a luxury.
A file doesn’t have Thursdays. It gets read before the first word, at the same standard, whether you’re sharp or finished. That’s the entire advantage, and it’s why the layer beats discipline over any period long enough to matter.
The layer also gets better in a way a habit can’t. Every time something slips through, the fix is one line added to a file, and that line is then load-bearing forever. Nothing you learn about your own voice has to be learned twice.
Write the list of phrases you’ll never publish. It’s the shortest part of the layer and it does the most work.
Build Your Content Machine. A free 5-part email course, one email per layer. Day 1 is this one, in practical form. Start the free course →
