The machine writes the text. You write the rules
Hand a neural network a brief and a minute later you have ten versions of the text, a set of images and a finished video. A week of a creative team's work now takes the length of one coffee. Every role in that scene has moved except one.
By Andrey Belokrylov · October 3, 2026 · 9 min read

Hand a neural network a brief and a minute later you have ten versions of the text, a set of images and a finished video. What used to take a team and a week of work now takes the length of one coffee. The executor, the creator who drew the banner or cut the video to someone else’s idea, is losing their indispensability fast.
One role has not moved an inch. For a company selling into Russia it turns out to be the whole game.
What the network already does, and what it still does not
A generative network does very well at anything you can describe in a brief. Describe the tone, the format, the length, the style, the references, and you get a result inside those limits. It is a fast, tireless executor ready to work in any genre, including in Russian.
Now look at the scene itself. For the network to generate anything, someone has to set the limits. Someone decides what the tone should be. Someone determines who the text is for, what job it does, which narrative it lives in. Someone writes the rule the generation will follow.
That part the network does not do for you. It generates content inside rules superbly, and it does not generate the rules. The rule arrives from outside, from the person who designs the communication. From the communicator.
Narrative design becomes a system of rules
Look at what follows for the profession. The copywriter’s and designer’s work used to be largely manual: sit down and write this text, draw this frame. That manual part is now handed to the machine more and more often.
The work that stays with a person rises one level. It is no longer “write the text” but “define the system by which the text will be written”. Not one artefact, but the rules from which a thousand artefacts will be born.
Narrative design is exactly this. First you define who your narrator is, what function they serve, whom they address, what world they act in, what voice they speak with. You assemble the frame: rules of delivery, the boundary of what is allowed, the logic by which the story develops. Only then, inside that frame, does generation start, machine or manual.
Value shifts from execution to intent. From the one who holds the pen to the one who holds the rule.
Why the Russian market makes this sharper
In a home market a company usually has its rules, written down or not. Years of talking to the same customers leave an implicit voice, and a model fed a few old campaigns picks it up well enough. When the same company turns to Russia, that implicit voice is gone, and what replaces it is usually a translation.
The pattern repeats across foreign accounts. The English site is written by people who know the buyer. The Russian site is produced by a model from a brief that says “translate and adapt”. The output is correct Russian that belongs to nobody: the jokes land the way English jokes land, the directness is the English level, the trust signals are the ones an American reader responds to. A Russian buyer recognises a foreign company speaking through a machine. They are not wrong.
The fault is not the model. Whether the generation runs through a Western model or a Russian one such as YandexGPT or GigaChat, the result is only as specific as the rules it received. “Adapt for Russia” is not a rule. A rule looks like this: the narrator is an engineer, not a sales manager; the reader is a procurement lead at a regional manufacturer who has been burned by an importer before; the voice is dry and concrete; we never promise delivery dates we do not control. Give any model that frame and the Russian text stops sounding translated. Give it none and the best model produces a fluent nothing.
Two further Russian specifics make the rule question less optional than in the West. The first is legal. Russian advertising law puts responsibility for advertising content on the advertiser and has no category of author called a neural network. If the model invented a certification claim, a discount that does not exist or a comparison with a competitor, the regulator’s letter arrives at your address. The rules are the only place in the chain where someone decides in advance what the text is not allowed to say. The second is how Yandex treats generated volume. Yandex Webmaster flags pages as low-value, and a site stuffed with machine text that says nothing specific is a candidate for that label. Rules that force specificity are also what keeps the pages in the index.
For now, a person holds the chaos back
The balance everything rests on today is simple. The network generates without limit, but a person sets the rules, and that person is what stops the stream from falling apart. The rule works like a dam: it turns an endless ability to produce anything at all into a meaningful, directed flow.
While the dam stands, everything is under control. The machine produces, the person directs. The chaos of generation is held by the frame of intent.
This is why brand strategy for a Russian launch is not a design deliverable. It is the dam. Everything the model later writes for the Russian market, from landing pages to Telegram posts to ad copy, flows through the same rules, and if they were never defined for this audience, no volume of generation will fix it.
What happens when networks start writing the rules
Now imagine the next step. What happens when networks can produce rules as well as follow them? When the system itself decides what the tone should be, who the text is for, which narrative it lives in?
That is not relief. That is chaos amplified. Right now the stream is held together by the fact that the source of rules is a person, someone with intent, responsibility and taste. Remove that source and rules begin to be generated as endlessly and uncontrollably as content is generated now. There will be as many frames as there are texts, and they will stop being frames.
When rules multiply with nothing underneath them, there is nothing left to orient by. Everything is possible, everything is generated, and so everything is equally empty, because everything is equally plausible. That is not a world of order. It is a world where order became so abundant it lost its value. For a foreign company in Russia that future arrives early, because the temptation to let the model decide what “Russian tone” means is strongest where nobody on the team can check.
The only anchor left: the customer’s authenticity
Here is what remains when frames multiply without limit. The anchor is no longer the producer of content and no longer the system that generates it. It is the person who consumes it, and more precisely their authenticity.
Authenticity here does not mean the genuineness of a thing. It means a person’s faithfulness to their own story. They have their own line of life, their own attachments, their own “this is mine and this is not”. When everything outside is generated and any rule is as good as any other, the only firm ground left is internal: does this fit me or not, does it belong in my story or is it foreign to it.
In a world where a machine can say anything convincingly and inside any frame, a person will hold on not to the brightest message but to the one that is most their own. Not to what was generated best, but to what resonates with their own script.
For a company selling into Russia from abroad this is the practical meaning of “know your customer”. The Russian buyer’s script is not a translation of the Western buyer’s script. What counts as a credible promise, what counts as pushy, what a company is expected to say about itself and to leave unsaid, all of it is part of a story the buyer already has. You cannot generate your way into it. You can only find out what it is, which is why customer research with real Russian buyers comes before any rule is written. A rule built on an assumed audience is a more elegant way of guessing.
What to do if you build communication now
The practical conclusion is simple, though not easy. Do not compete with the machine where it is already stronger: speed and volume of execution. Move to the level it has not yet claimed: intent, rule, narrative.
For a company selling in Russia, that level has a concrete shape:
- Write the rules for the Russian audience separately. Do not inherit them from the English brand book by translation. Narrator, reader, voice, limits, the claims you will never make.
- Find out the buyer’s own story before writing the rules. Talk to Russian customers, read how they describe the problem, look at what they distrust. The rules come out of that, not out of a persona slide.
- Put a person who reads Russian between the model and the publish button. To check that the output stays inside the rules and says nothing the advertiser would have to answer for.
- Treat the model as interchangeable and the rules as the asset. Models change every year. A well-defined frame outlives all of them.
- Prefer fewer, more specific texts to more generic ones. Yandex and the reader both punish volume that says nothing, and generated volume is the easiest thing to produce.
Learn not to write the text but to design the system the text follows from. Define the function, the audience, the narrator’s voice, the logic of the story, and hand the generation to whoever is faster. Your value is in the frame, not in the filling.
And prepare for a world where even frames are produced automatically. There the winner is not the one who shouts loudest but the one who hit the customer’s authenticity, their own story, which nothing generated. That is the one thing a machine can neither fake nor cancel.
The machine has taken the pen. Soon it will reach for the decision about what to write with it. One line will remain that no generation crosses: a living person with their own, uninvented story. In Russia that person is a buyer your company does not yet know well, which is why the rules cannot be outsourced. Whoever understands this earlier will not fight the stream. They will stand where the stream breaks, at the authenticity of the person the whole thing was for.
Frequently asked questions
If generative AI writes Russian copy well, why do we still need a strategist?
Because the network works inside a brief and does not write the brief. Someone has to decide who the narrator is, which Russian buyer the text addresses, what voice it uses and where the limits are. A model given no rules produces fluent Russian that belongs to nobody. A strategist's job has moved from writing the text to designing the system the text comes out of.
Can we just translate our English brand voice into Russian with an LLM?
You can translate the words. The rules behind them will not survive the trip untouched. An English voice built for a Western buyer carries assumptions about humour, directness and trust that a Russian reader decodes differently. The rules have to be redefined for the Russian audience first, and only then handed to the model, whichever one it is.
Does it matter whether we generate Russian text with a Western model or a Russian one like YandexGPT or GigaChat?
Less than people expect, as long as the rules are yours. Any of them will produce plausible Russian inside a well-written frame and generic Russian inside a vague one. The choice of model is a tooling decision. The choice of narrator, audience, voice and limits is a strategic decision, and it is the one that shows up in whether a Russian customer recognises anything of themselves in the text.
Who is liable in Russia for what the model wrote in an ad?
The advertiser. Russian advertising law does not know a category of author called a neural network. Whatever went out under your brand, you answer for it, including claims the model invented. That is another reason the rules must sit with a person on your side: they are the only point in the chain where responsibility can actually live.
Sources
- Russian version of this article on belokrylovo.ru: Машина напишет текст. Правила напишете вы