Your Voice Guide Is a System Prompt Now
When AI drafts the first version, brand consistency depends on rules, examples, and one operational source of truth.Tom Orbach spent his July 30 issue attacking what he calls the brand-archetype industrial complex, the world of “12 astrology cards” and mood boards, and reporting what he found when…
When AI drafts the first version, brand consistency depends on rules, examples, and one operational source of truth.Tom Orbach spent his July 30 issue attacking what he calls the brand-archetype industrial complex, the world of “12 astrology cards” and mood boards, and reporting what he found when he collected the internal voice guides actually used at billion-dollar companies.The working documents look nothing like the archetype decks. They are short, practical messaging guides built for employee onboarding: tone specs per channel covering email, social, website, sales, and in-product copy, stocked with real examples from companies like Amazon, Wiz, Shopify, and Stripe.He turned the pattern into a fill-in-the-blank template, formatted deliberately so it can be pasted into AI tools to generate on-brand copy, and he states the quality bar plainly: an employee should be able to absorb the voice in ten minutes.Buried in the packaging is the insight I think deserves its own essay, and Orbach says it almost in passing: a brand voice document is an operations tool, and its new primary reader is an LLM. I want to take that sentence as seriously as it deserves, because if it is true, and I believe it is, then most companies’ brand documentation is written for a reader who no longer does the reading, in a format the actual reader cannot use.The astrology-card era, and why it lastedLet me start with the corpse Orbach is kicking, because it earned the kicking. For two decades, the standard deliverable of a brand voice engagement has been a document built for the meeting where it gets presented.The archetype card: your brand is The Sage, or perhaps The Explorer. The persona exercise: if our brand were a person, she would drink oat-milk cortados and drive a Volvo. The mood board of borrowed photography. And the adjective slide, words like bold and approachable, chosen in a workshop where nobody was allowed to propose timid and standoffish, which tells you exactly how much information the chosen words carry.I say this as someone who has sat through the workshops and, in weaker moments, enjoyed them. There is real pleasure in an afternoon of talking about what your company would be like as a dinner guest. The pleasure was the product.The genre persisted because it served everyone in the room except the writer, who would need it later. It gave the agency a deliverable with high perceived depth and no falsifiable content. It gave the leadership team an afternoon of pleasant self-contemplation. It photographed well in the brand book.What it never did was answer the question a person drafting an actual email at 4 p.m. has, which is: can I say this, or not? I have watched writers hold an adjective slide in one hand and a draft in the other, and the slide simply does not adjudicate. Bold and approachable permits everything and forbids nothing. The archetype deck was unfalsifiable, and unfalsifiable documents cannot govern anything.The companies Orbach studied apparently figured this out years ago, at least internally. The guides that run real voices are boring the way an airline checklist is boring: channel by channel and rule by rule, with the examples doing the teaching. That gap between the public genre and the private practice is itself a finding. The performative document and the operational document split, and the operational one is the only one with a job.The reader changed before the documents didNow the part that turns a craft observation into a strategic one. Count where copy volume actually comes from in a 2026 marketing team. The emails, the social variants, the landing-page iterations, the ad permutations, the help-center rewrites, the localization passes.In most teams I see, the first draft of most of that is now machine-generated, with humans editing, approving, and occasionally rescuing. Whatever the official policy says, the distribution of first-draft authorship has flipped, and it is not flipping back.Follow the consequence. If a model drafts most of the words, then the single document with the most influence over your public voice is whatever gets pasted into the model’s context before generation. Not the beautiful brand book on the shared drive.The paste. And in companies with no operational voice guide, the paste is improvised: each employee’s personal recollection of the brand, reconstructed from memory into an ad-hoc prompt, differently every time, in every corner of the org. The voice you ship is the average of a hundred private impressions of the voice, sampled through a model. Nobody decided this. It happened while the brand book was being admired.If the flip sounds abstract, walk through one ordinary Tuesday. A lifecycle marketer needs a win-back email. She opens her drafting tool, types a two-line instruction, and pastes whatever brand context she keeps in a personal note, a paragraph she wrote herself eight months ago, half-remembered from a workshop.Down the hall, a product marketer generates launch copy from a different personal paragraph. Support is generating macro rewrites from none at all. Every one of those outputs is on-brand relative to its paste and off-brand relative to each other, and the aggregate, which is what customers actually experience, has no author. Multiply by a year of Tuesdays. That is the current voice governance of most mid-sized SaaS companies, described without exaggeration.Orbach’s template formats the guide for pasting into AI tools, and that formatting choice is the tell that he understands the shift. The voice guide has become, functionally, a system prompt: the standing instruction set that constrains a generator. Once you name it that way, the quality criteria rewrite themselves. A system prompt is not judged by how well it captures the brand’s soul in a workshop. It is judged by whether the output obeys it, measurably, run after run.Adjectives are unenforceable; examples are lawHere is where the astrology cards go from useless to actively incompatible. A model handed “confident but humble, playful but professional” does what humans do with it, which is whatever it was already going to do, with a light stylistic seasoning.Abstract adjectives barely constrain generation, because they barely mean anything at the level where generation happens. What constrains a model, and this maps directly onto what the billion-dollar guides contain, is a different set of materials entirely.Explicit rules, stated as rules. Sentence-length ranges, punctuation prohibitions, second person versus first, how the product is named on first mention. Rules a checker could verify, which means rules a generator can follow. Channel-specific specs, because tone that is right in a nurture email is wrong in an error message, and a guide that speaks per channel matches how both the employee and the model actually encounter the work, one artifact at a time.And above everything, examples in volume: real sentences in the voice, real sentences out of it, before-and-after pairs. Models are imitation engines before they are instruction-followers. Five strong paired examples move an output distribution further than a page of description, and every practitioner who has iterated on a serious prompt knows it.The working guides are example-dense because the human junior writer, it turns out, was also an imitation engine all along; no CMO needed to anticipate LLMs to end up writing well for them. What onboarded her onboards the model. Orbach’s ten-minute absorption test survives the reader swap almost untouched.The negative space matters just as much. A banlist, the words and constructions the brand never uses, is close to inert for human readers, who skim it and forget, and close to binding for a model with it in context.If your guide has no “never” section, it is missing the cheapest control surface it can have. Write down what the brand does not say, specifically, and the machine will actually not say it, which is more than can be claimed for most human teams.Run it like infrastructure, because now it isOnce the guide is a system prompt, it inherits the operational disciplines of one, and this is where most brand teams will need habits they have never had. Test it: paste the guide and generate across your channels, then judge the output against your own ear.Where the output drifts, edit the guide rather than the copy, adding the example or the rule that would have prevented the drift, so the correction compounds across every future generation instead of dying in one document. This loop — output review feeding spec revision — is how engineers maintain prompts, and it is now how brands maintain voices.Version it, date it, give it one owner, and kill the copies.A guide that lives in six stale variants across personal drives is now six divergent system prompts shipping six divergent voices at machine speed. The consolidation is dull governance work, and it has quietly become brand-critical in a way it never was when humans, with their slow drafting and their shared office osmosis, did the writing. Voice drift used to spread at the speed of hiring.It now spreads at the speed of generation, which is a good argument for treating the paste-able guide as the single source of truth and auditing what actually gets pasted, at least occasionally. I know of no team doing that audit systematically yet, which I suspect will read as ancient negligence in about three years.The same infrastructure lens settles the length argument that brand purists will raise. Yes, a short guide flattens some richness. A guide competes for a model’s limited attention exactly as it competes for an employee’s, and an eighty-page book loses that competition to its own appendix.The discipline of the ten-minute document, in front of both audiences, is the discipline of deciding what actually matters about your voice and cutting the rest. Compression, here, is not a loss of depth. Compression is the proof you knew what the voice was.Building yours, field by fieldFor the team starting from an archetype deck and a guilty conscience, here is how I would sequence the rebuild, using the shape Orbach extracted from the real guides. Start with the channel grid, because it forces concreteness immediately.Five columns for wherever your brand actually speaks, on the pattern of the email, social, website, sales, and in-product split the working guides use, and for each one, a few sentences of tone direction plus two real examples of your best existing copy in that channel.Harvesting the examples is the valuable labor: you are forced to decide which of your past sentences actually sound like you, which is a sharper voice-definition exercise than any adjective workshop, because it grades artifacts instead of aspirations.Then write the never section before the always section. Prohibitions come to mind faster; they are more enforceable, and they carry more identity than teams expect. The specific jokes you do not make, the punctuation you do not touch, the way you never talk about competitors, the corporate phrases that would make your founder wince. Ten crisp nevers constrain a generator more usefully than a page of personality prose.Then the naming and mechanics block, the unglamorous facts every writer and every model needs: how the product and features are capitalized, how customers are referred to, formatting conventions, reading-level target. This is the section nobody debates, and everybody violates, and having it in the paste ends the violations at their source.Last, run the acceptance test, both versions of it. Hand the document to the newest person on the team and time their absorption against the ten-minute bar. Then paste it into your drafting tool, generate one artifact per channel, and read the outputs cold.Wherever either reader produces something that made you flinch, the edit goes to the document rather than the reader. Two afternoons of this loop produce a guide that outperforms most six-figure brand engagements on the only test that matters, which is what ships.Where the philosophy still belongsI want to end on the honest counterweight, because it would be easy to read this essay as “replace the brand strategists with a prompt file,” and that is not the claim. The astrology cards were bad answers, but the question they were mangling is real: what does this company believe, and what should it sound like because of who it is?A spec cannot generate its own content. Someone still has to decide that the brand is blunt where competitors are ingratiating, or warm where the category is clinical, and that decision is positioning work, upstream of any document format, unautomatable in exactly the way choosing is always unautomatable.The order of operations is the whole resolution. Decide the voice as a strategic act. Then encode it as an operational spec, example-rich, rule-explicit, channel-aware, short enough to absorb in ten minutes, formatted for the reader that now drafts most of your words.The companies Orbach studied are not less thoughtful about voice than the archetype crowd. They are thoughtful in a format that survives contact with Monday morning, and now, without ever intending it, in the one format that survives contact with the machines. The mood board was for the meeting.The guide is for the writer, and the writer, most hours of most days, is a model with your document in its context window, doing exactly what you wrote down, which has become the strongest reason in the history of brand documentation to write down the right things.This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!Your Voice Guide Is a System Prompt Now was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI