How I write with AI every day and it still sounds like me

The reason most AI writing sounds like AI writing isn’t the model. It’s that people give it nothing to work from, so it writes the average of everything — and the average of everything is precisely what “sounds like AI” means.

I use it constantly, for client copy and for my own. What stops it flattening my voice isn’t a clever prompt. It’s several thousand words of written-down decisions that go in before I ask for anything, plus a hard rule about which part of the process it’s allowed to touch at all.

This post is that system. It’s a real one I maintain and use daily, not a framework invented for the occasion.

Why does AI writing sound like AI writing?

A language model produces the most probable next thing. That’s not a criticism, it’s the mechanism. Given no constraints, the most probable business writing in existence is competent, symmetrical, slightly over-explained and completely anonymous.

You know the shape the moment you see it. Three points, because it’s always three. A sentence that sets up a contrast and then resolves it a little too neatly. A closing paragraph summarising what you’ve just read. Nothing wrong with any of it. Nothing anybody would recognise as a person, either.

You can’t prompt your way out of this with adjectives. “Write in a punchy, conversational tone” produces the average of punchy and conversational, which is a LinkedIn post about resilience. The fix isn’t a better description of the voice. It’s more decisions made in advance.

The four documents I load before writing anything

In order of how much difference they make.

1. A document about what I write about

Not the voice — the substance. What the whole operation is about, what it isn’t, who it’s for, and the specific topics that recur. Mine names five pillars with a rough share of output for each, two audiences with different reasons for reading, and a short, deliberate out-of-scope list.

That out-of-scope list is the most useful page in the whole set. “No income promises. No recipes for things I haven’t done. No criticising named creators — criticise the pattern, never the person.” Those are the constraints that stop a draft drifting towards whatever happens to rank.

2. A document about who I am in public

The role, the positioning, the disposition underneath the writing. Mine includes a table of Am / Am Not pairs, which does more work than any quantity of tone description:

Pairs work far better than single adjectives, because every quality has a failure mode living next door to it. “Direct” on its own drifts into rude by the third paragraph. “Direct, and here’s the specific thing it must never become” holds all the way down the page.

3. A set of copy rules

The mechanical layer. British English throughout. First person singular — I build, I sell, I got stuck — never “we,” because there is no we and a buyer catches that immediately. “One-person studio,” never “one-man show.” Specific over impressive: a real tool, a real week, a real number, because specificity is what makes authority credible and adjectives are what make it suspicious.

These accumulate from corrections. Every time I fix something in a draft, the fix becomes a rule, so I never have to fix it twice.

4. The copy that’s already settled

The highest-leverage part, and the one nearly everybody misses. Any line that’s been decided — a hero paragraph, a tagline, a section header, the standard description of what I sell — gets stored verbatim.

Without this, every draft quietly re-paraphrases things you already got right. You end up with four slightly different versions of your own positioning across four pages, none of them the good one. Storing the settled version and instructing “reuse, don’t re-derive” kills that off entirely.

The rule that matters more than all four documents

It never gets the first thought, and it never gets the last pass.

The first thought is mine, because the idea has to come from having done the work. If I outsource what to say, there’s nothing left worth reading — the entire value of a maker writing about making is that it actually happened to them.

The last pass is mine, because that’s where the specific detail goes in: the thing that annoyed me on Tuesday, the tool that broke, the number that came out worse than I expected. None of that exists inside any model.

The middle is where it earns its keep. Restructuring an argument I’ve made badly. Telling me the third section is doing the same job as the first. Producing four openings so I can see which angle is genuinely strongest. Being a reader who has read the whole thing and will tell me where it sags.

That last one is the real use case, and it’s badly underrated. Editing is the expensive part of writing alone, because you cannot see your own draft properly an hour after writing it. Having something that can is worth considerably more than having something that can generate.

Where it still goes wrong

It would be dishonest to write this without the failure list, so here it is.

It reaches for symmetry. Give it three points where the third is thinner, and it’ll inflate the third to match rather than tell me it’s weak. I have to ask directly: which of these is doing the least work?

It hedges. Left alone it’ll soften a claim I meant to make sharply. “This can sometimes be an effective approach for many teams” instead of “do this.” I strip modifiers on every pass, and there are always some left.

It writes closing summaries nobody needs. A final paragraph restating the post I’ve just written. I delete it roughly every single time.

It cannot tell when something is actually funny. It knows the shape of a joke. Given permission to be funny, it produces something joke-shaped that isn’t one. My rule is that if the joke isn’t already there, don’t reach for it — and this is the rule I most often have to enforce by hand.

It drifts back to the mean over long documents. The first thousand words hold the voice well. By the third thousand it’s sliding towards generic. Long pieces need re-anchoring partway through, or writing in sections.

The documents rot if you don’t update them. This is the real maintenance cost. A rule set reflecting what you thought six months ago produces confidently wrong copy. Mine are versioned with changelogs, and the discipline is to add the rule in the same session the decision gets made — not to schedule a review that never actually happens.

Does this save time?

Not on the first piece, no. Writing down what you write about, who you are and what you sound like takes a full day if you do it properly, and it’s genuinely hard work — most of it is making decisions you’ve been comfortably avoiding.

The return comes later, and it isn’t mainly speed. It’s consistency. Every bio across eleven channels says a compatible thing. The homepage and the sales email describe the same offer in the same words. Nothing contradicts anything else.

That’s a job I used to do badly by hand, because keeping eleven profiles aligned in your head is not something a head can do. The documents don’t just make the AI better. They made me better, because writing them forced decisions I’d been leaving deliberately fuzzy.

Which is the actual argument, and it’s slightly annoying: the useful part wasn’t the tool. It was that using the tool properly required me to write down what I believe.

The system, in short

Common questions

Which model do you use? It matters far less than the brief. I’ve run this set through several and the difference between models is smaller than the difference between a loaded brief and an empty one. Pick whichever you’re happy paying for and spend the saved decision on the documents.

How long does the setup take? A full day to write the first version properly, then a few minutes a week to maintain. The maintenance is the bit people skip, and it’s the bit that makes it keep working.

Isn’t this just prompt engineering with extra steps? It’s the opposite, really. Prompt engineering tries to describe the output in the request. This moves the decisions upstream into stable documents so the request can be short and boring. My actual prompts are usually one line.

Will readers be able to tell? They’ll tell if you skip the last pass. Nothing else in the process gives it away as reliably as the absence of specific, lived detail — the number that disappointed you, the Tuesday it broke.

Does this work for client copy as well as your own? Yes, and it’s arguably more valuable there. Each client gets their own version of the four documents, which also means the positioning work gets done properly at the start rather than being improvised page by page.

---

I write about building, selling and shipping as a one-person studio at tuomo.co — including the parts that didn’t work.