Ask a general model for a post about SaaS pricing, then ask it for one about marathon training. Read them side by side. The topics are different. The shape is identical: an opening that restates the question, three points, a reversal in the middle, a closing line that reaches for profundity.
That shape is what people recognise. Not the vocabulary. The structure. And it comes from one place: a single prompt doing duty for every subject on earth.
Why one prompt cannot cover every subject
A prompt that must work for fitness, philosophy, crypto and B2B sales can only contain instructions true of all four. That is a very small set. It reduces to generic writing advice: be specific, hook the reader, keep it tight, which is exactly the advice that produces generic writing.
Everything that would make a post good in a particular field is field-specific, and therefore excluded:
- What counts as evidence. In crypto, a number and a timeframe. In philosophy, a distinction. In fitness, what happened to a body over months. Interchanging them makes the writing wrong in a way readers feel before they can name.
- What the reader already knows. Explaining what churn is to a room of founders reads as condescension. Not explaining it to a general audience reads as jargon. One prompt has to pick a lane and be wrong for someone.
- What has been said to death. Every field has phrases that were insight three years ago and are now wallpaper. They are completely different phrases in each field, and a general prompt cannot ban them all. The list would be longer than the prompt.
The third one is where most AI tells live. Not in exotic phrasing, but in a field's exhausted vocabulary, deployed with total confidence by something that has no idea it is a cliché.
What we did instead
We split the system into 48 engines, one per niche. Each carries its own audience definition, its own content pillars, its own structural rules, and its own list of phrases it will never produce.
The banned lists are the part that took the longest and the part we do not publish. They are specific. The phrase, not the category, and they were assembled by reading a lot of bad posts. Publishing them would hand over the only part of this that is genuinely hard to rebuild.
The audience definitions and the topic pillars we are happy to show; they are positioning, not method. You can read them on the niche pages, along with a real sample of each engine's output.
The mistake that made everyone in a niche sound the same
Here is the part worth reading if you are building something similar, because it is the opposite of what you would expect.
Every one of our 48 engines contained a worked example: a block labelled EXAMPLE OF CORRECT OUTPUT FORMAT, showing a well-formed post in that niche. Standard few-shot prompting. It works, outputs came back correctly structured.
It worked too well. The model was not just copying the format of the example. It was copying its opening pattern. Which meant every user writing in a given niche received a post built on the same skeleton as every other user in that niche.
We had fixed sounding-like-AI and replaced it with sounding-like-each-other. From a reader's point of view that is the same failure: your post looks like something they have seen before.
The fix was not to delete the examples, since without them formatting degrades. It was to say centrally and explicitly that the example demonstrates format only, never sentence patterns. Output openings diversified immediately.
A few-shot example teaches more than you asked it to. Whatever is consistent across your examples is what the model learns, including the parts you did not intend as the lesson.
There is a second version of this problem we hit later. When a user has a voice profile of their own, the niche engine's voice section is no longer a helpful default. It is a competitor. Two sets of style instructions, and the more forceful one wins. So when a profile is present, we now strip the engine's voice section and its worked example entirely, keeping only the domain knowledge. That change is described in more detail here.
The honest limitation
Per-niche engines fix the part of sameness that comes from your tool. They do not fix the part that comes from the underlying model.
We removed our worked examples and outputs still opened by restating the topic, across every niche, because that is a habit of the model itself rather than of our prompt. Specialisation gets you a long way and then you meet a floor that no amount of instruction moves.
That floor is why we score every draft before showing it to you: hook strength, readability, length, and density of AI tells, rather than assuming a good prompt produced a good post. The prompt is a bet. The score is a measurement.
What this means if you are choosing a tool
A reasonable test, and it takes about four minutes:
- Generate a post about your actual subject.
- Generate one about something completely unrelated. A different industry, a different register.
- Read the two openings and the two closings next to each other.
If they are structurally identical, you are looking at one prompt with a topic slotted in, and your readers will eventually see the same thing. Do it to us as well. That is the honest way to run the comparison.