When one team writes for three brands, the three often end up sounding the same. The group website, the consumer brand and the B2B subsidiary drift towards the same average voice: polite and a little vague. Now that some of the writing goes through an AI model, it is even more obvious. Without precise instructions, the model writes like everyone else. Editorial guidelines for each brand are there to stop that.
I write them for every site my content machine runs on. The engine is the same everywhere; what changes is a voice file and a rules file. Here is how I build them, what goes in them, and how I check that they are followed, by a writer as much as by an AI.
01Editorial line, style guide, brief: what each one does
The three terms are often used as if they meant the same thing. The way I use them, they cover three different levels, and keeping them apart makes the work much simpler.
The editorial line says what the brand talks about, to whom, and in what voice. It rarely changes. The style guide turns that voice into writing rules: words to avoid, how to cite a source, paragraph length, how to address the reader. It evolves as reviews come in. The editorial brief applies both to one specific piece: this keyword, this reader, this promise, these sources.
The usual trap is a single, very long document that mixes all three. Nobody rereads it, and an AI model given the whole thing mostly picks up the noise. I have described elsewhere how I split up the writing work in my AI content machine with human review; the editorial guidelines are its first building block.
02Why one set of guidelines per brand, rather than one for the group
Group guidelines give a shared personality. They do not say who reads each brand, or what each brand is allowed to claim. Yet that is exactly what tells two sites from the same group apart in a reader's eyes.
Google's documentation asks the question plainly: does your site have a primary purpose or focus, and an audience that would find your content useful if they came to you directly? It also warns against producing lots of content on many different topics (Google Search Central, helpful, people-first content). Guidelines per brand force you to answer both questions for each site, instead of leaving them floating at group level.
The same goes for different voices within one company. At Moonworkers, my own company, the voice depends on who is speaking. The founder's profile speaks in the first person, with an opinion. The company page stays factual, in the third person. Those are two sets of guidelines, not one set with exceptions. It is the same logic as a multi-brand Figma design system: a shared foundation, and what makes each brand different written down separately.
03The five sections of editorial guidelines that fit on one page
I hold myself to one page. Beyond that, the document stops being read, by a person or by a model. These are the sections I fill in, in this order.
1. Who is speaking, and to whom
One sentence on the speaker (a person, a team, a company) and one sentence on the reader, in their real situation. Not "SMEs", but "a founder who has to choose a payroll tool without an in-house accountant". The more precise the reader, the easier the tone is to set.
2. What the brand is allowed to claim
This is the section most often forgotten, and the one that protects the most. It lists the facts the brand can rely on: what it has done, what it has measured, what went wrong. For my own blog, it is a closed list. If an article needs an anecdote that is not on it, the article is not written: it waits for my decision. An AI model will happily invent a believable story, and a false story does more damage than a text with no story at all.
3. Tone, in words you can test
Nielsen Norman Group suggests describing tone on four dimensions: formal or casual, serious or funny, respectful or irreverent, matter-of-fact or enthusiastic. The article advises sticking to a handful of tone words and adding "anti-tone" words, the tones you want to avoid (Nielsen Norman Group, the four dimensions of tone of voice). The video version of the method goes further: start from existing content that already sounds right, then keep three words to aim for and three to avoid (NNgroup, Establishing Tone of Voice).
I also like the way GOV.UK writes its tone: each quality is bounded by its excess. "Brisk, but not terse"; "serious but not pompous" (GOV.UK, Use the right tone). I find that form very useful to borrow, because it tells the writer where to stop.
4. Yes and no examples
Two or three sentences that sound right, two or three that sound wrong, with a line saying why. For this blog, the reference "no" sentence is a pompous introduction that lists years in the trade and a "unique approach" without showing anything. An example gets across in ten seconds what a paragraph of instructions cannot.
5. What the brand never says
Excluded words, promises and topics. Here: no promised results, no self-promotion, no named clients. This section is the easiest to check, so it is the easiest to enforce.
04Making the guidelines usable by a writer or an AI
Guidelines you "keep in mind" do not get applied. I make them usable in three ways.
First, they are handed over as they are, every time. In my machine, the model reads the voice file before writing, then the approved brief, then the yes and no examples. I describe the exact order in my piece on the article writing prompt.
Anthropic's documentation confirms that examples are among the most reliable ways to steer the tone and format of a response, and recommends 3 to 5 relevant, varied examples kept separate from the instructions (Anthropic, prompting guide, use examples effectively). The same page notes that explaining the reason behind a rule helps the model generalise it, so I write the why next to every ban.
Second, the guidelines are kept apart from the mechanical rules. Tone and angle are a matter of judgement; a list of banned words can be checked. I put the latter in a configuration file read by a tool, and keep the former in the voice file. A rule repeated in a prompt gets forgotten; a rule in the config blocks.
Third, they grow from reviews. When a reviewer corrects the same thing twice, I do not fix it a third time by hand: I write the rule. I tell that story in my piece on reviewing AI-generated content.
Here, in short, is what the first page looks like for a fictional brand:
``` Who speaks: the product team of a scheduling tool for tradespeople. Who reads: a tradesperson with three to ten staff, in the evening, on their phone. Can claim: what the tool does, tested; nothing about clients. Tone: simple without being chummy; precise without being technical. Avoid: excitable, salesy, condescending. Yes: "Move a job, and the team's schedule follows." No: "A unique solution that transforms your day-to-day." Never: a time-saving figure without a source. ```
05Checking that the guidelines are followed
Guidelines nobody checks end up as decoration. I check at two levels.
The first is automatic. Before publication, a check rereads the text against the style guide: banned words, em dashes, paragraphs that are too long, first person present or absent depending on the brand, figures without a source. Red, the text does not go out. Amber, it goes out only if I can justify the point in writing. This check does not judge tone; it catches what can be counted.
The second is human, and it is the only one that tells you whether the tone is right. Nielsen Norman Group suggests having people read extracts with no logo or colours, then asking them to pick words to describe them (NNgroup, Establishing Tone of Voice). If you were aiming for "understated" and people say "cold", the guidelines are not right yet. This test does not need to run on every article: it is useful when a brand starts out, or when doubt creeps in.
The limit is real: all this takes time at the start. For a brand that publishes three articles a year, a one-page voice file and a careful reviewer are enough, with no tooling. Tooling pays off when volume grows, or when several people and a model write for the same brand.
06Where to start
If you manage several brands, I would start on Monday with a simple exercise. For each one, reread ten recent pieces. Keep the two or three that sound most right, and write one line on why. Note the one you wish you had not published, too. With that, you already have the reader, three tone words and your first yes and no examples.
Then put it all on one page per brand, and hand that page, as it is, to the next writer or the next prompt. You will quickly see what is missing.
This is the kind of work I take on in short engagements, from the voice page to the machine that applies it: you will find the details in my services and examples in the portfolio. If you would like to talk it through, we can.