If you have used AI for writing, you will have tried this. The output comes back stilted and slightly obnoxious, so you type: write more naturally.
And it does not help.
The reason is worth understanding, because it explains why most attempts to fix this fail. The model does not know it is writing unnaturally. Its style is not a mistake it is making. It is simply how it writes, and asking it to stop being itself is not an instruction it can follow.
I have written before about the tension between writing for machines and writing for humans. This post is the practical companion: how to actually specify a voice, in enough detail that it works.
First, why this is getting harder rather than easier
A reasonable assumption a couple of years ago was that as models improved, the writing would become more natural. It has not really gone that way.
Part of the reason is what the labs are optimising for. An enormous amount of effort is currently going into making these things better at software engineering, which is a different sort of writing altogether. Terse, dense, context assumed rather than stated, jargon used freely because the reader is expected to know it.
You can feel that in the prose. Sentences that read like they have been compressed. Important context left out because it was obvious to the writer. A general tendency towards efficiency over warmth.
So this is not a problem that solves itself by waiting. If anything you need to specify more now than you did.
Layer one: ban the obvious tells
The first layer is the deterministic stuff, and I have covered it in more detail elsewhere. Briefly: name the specific habits rather than describing the vibe.
No em dashes. No contrastive framing, the “it is not X, it is Y” construction. No empty intensifiers. And most people end up with a list of overused words and phrases they never want to see.
One thing worth knowing about those lists: they go stale. The tells differ between models and they change as the models change. A ban list you wrote a year ago is probably policing habits that have gone and missing the ones that arrived since. Treat it as something you maintain rather than something you finish.
This layer gets you a long way and then stops working entirely. At some point you have to stop telling it what not to do and start telling it how to write.
Before you specify anything: which writing?
A mistake almost everybody makes, including me for a while.
You do not have one writing style. The way you write a blog post is not how you write a sales page, and neither is how you write a social post. Trying to define one voice for all of it produces guidance so general it does nothing.
So split it. Decide what applies to everything you write, and what applies only to a particular type.
Universal is usually the character stuff: how formal you are, whether you swear, what you would never claim. Category-specific is usually structural: length, how direct the opening is, whether you use headings, whether it ends in a call to action.
Getting that separation right before you write any guidance saves a great deal of rework.
Layer two: actually specify the voice
Now the important part, and the rule underneath all of it: the more specific, the better.
Vague guidance produces the worst output there is. Tell a model to “add more personality” and you will get personality, of a sort. It will be forced, it will be trying too hard, and it will be recognisably nobody.
Three things that genuinely help.
Score yourself on scales
Rather than describing your voice in adjectives, place it on scales. Four that I find cover most of it:
- Warm to authoritative
- Formal to conversational
- Straightforward to polite
- Playful to serious
The point is not to pick an end. It is almost always somewhere in between, and the position is the useful information. “Mostly conversational but not matey” is far more actionable than “friendly and professional.”
That third scale is the one people find hardest and the one that matters most. Straightforward means saying the thing. Polite means softening it. Most businesses claim they want straightforward and then flinch at the output, which is worth knowing about yourself before you brief anybody, human or otherwise.
Define who is writing
Give it a person. Not a brand, a person.
Who are they, what do they do, how long have they been doing it, what do they find irritating, what would they never say? The more concrete the person, the more consistent the output.
If your style is close to somebody well known, say so. That is an enormous amount of information conveyed in a few words, and models are good at using it.
Set a level
Specify who you are writing for and how plainly. A reading age of around twelve, or explain it as you would to somebody intelligent who does not know the subject. There are also formal standards for plain writing, such as Simplified Technical English, originally developed for aerospace manuals, which some people use as a reference point for stripping out flourish.
Whatever you choose, it is another constraint, and constraints are what produce consistency.
Layer three: examples, and how they go wrong
The final layer is showing it what good looks like. This is the most powerful and the easiest to get badly wrong.
The problem is that models are extremely good at spotting patterns, including patterns you did not intend to demonstrate.
If all five of your examples are written in flowing paragraphs, it concludes you never use lists.
If two of them contain the same phrase, it concludes that phrase is a signature and will use it constantly.
If two of them are about the same topic, it concludes a substantial share of your writing is about that topic and will steer everything in that direction.
None of that is the model being stupid. It is doing exactly what you would do if handed five samples and asked what they had in common.
Which means your examples must be as different from each other as possible, sharing only the thing you actually want copied.
Different lengths. Different formats. Different subjects. Different degrees of formality if your voice genuinely varies. The only constant should be the voice itself, because whatever else is constant will be copied too.
That takes proper effort. You have to go through what you have written, choose deliberately, and reject good samples for being too similar to ones you have already chosen. Which is tedious, and it is also the bit that makes the difference.
The honest limit
All three layers together get you most of the way. They will not get you all the way, and I would be lying if I said otherwise.
What closes the last bit of the gap is correction over time. Every time the output is not quite right, work out why, and turn that into another piece of guidance. Do that for a year and you have something genuinely specific to you.
There is no version of this where you write the perfect brief once and walk away. The guidance is a living document, and the work is in maintaining it.
Which is the same as briefing a new member of staff, really. You would not hand somebody a document on day one and expect never to correct them again.
Want content that actually sounds like your business? Get your free audit, or email hello@inkon.co.uk. I read every message myself.