In short: A brand voice guide for AI is a one to two page brief that tells the model who you write for, how you sound, how you address readers, which words you use and avoid, how long sentences and paragraphs are, how posts are formatted and which claims are off limits, plus two or three paragraphs of real text in your style. Write it in concrete do and don't pairs, not adjectives. Test it on one sample post, fix the guide (not the post) and repeat until two drafts in a row need only small edits.
Why AI text sounds generic
A language model writes the most likely next words for the instructions it gets. If the instructions only say "write a blog post about garden irrigation", the most likely result is the average of every irrigation article it has seen: a friendly opening, a list of benefits, a summary. That is the "sounds like ChatGPT" effect people complain about in forums.
Two things make it worse:
- Vague adjectives. "Friendly, professional, approachable" describes almost every company, so it changes almost nothing.
- A blank slate every time. Many tools treat each prompt on its own. If the voice lives in one person's head or in an old chat, every new post starts from zero, and the voice drifts.
The AI vendors say the same thing in their own documentation. Anthropic writes: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result" (Anthropic: prompting best practices (external link)). OpenAI compares a GPT model to "a junior coworker" that will "perform best with explicit instructions to create a specific output" (OpenAI: prompt engineering (external link)). A voice guide is those explicit instructions, written once and reused.
Brand voice and tone: what is the difference?
Voice is who you are, and it stays the same everywhere: your values, your attitude, your vocabulary. Tone is how that voice adapts to the situation: a how-to guide is calmer than a launch announcement, an apology is more careful than a tips post. A good guide sets the voice once and then names the tone for the content types you actually publish.
The parts of a voice guide an AI can follow
Each part answers a question the model would otherwise guess.
| Part | The question it answers | Write it as |
|---|---|---|
| Audience | Who reads this, and what do they already know? | One or two sentences about a real reader |
| Voice and tone | How do we sound? | Three to five do and don't pairs |
| Form of address | Do we say "you", "we", first names? (German: Sie or du) | One clear rule |
| Words to use | Which terms do our customers use? | Short list with the preferred word |
| Words to avoid | What do we never say? | Short list, with a replacement where possible |
| Sentences and paragraphs | How long, how simple? | A rule of thumb, for example "most sentences under 20 words" |
| Formatting | Headings, lists, bold, length? | Short rules per content type |
| Facts and claims | What may the AI state on its own? | Clear limits and what to do instead |
| Examples | What does good look like? | Two or three real paragraphs |
Audience
Describe one reader, not a market segment. "Homeowners in and around Bristol with a garden of 100 to 400 square metres, who do some work themselves but hire help for bigger jobs" gives the model something to work with. "Homeowners aged 30-65" does not. Add what the reader already knows and what they worry about, because that decides which terms you explain.
Voice and tone as do and don't pairs
Adjectives alone are too vague. Pair each one with what it means and what it does not mean:
- Practical: give the step, the tool and the time it takes. Not: general tips like "plan carefully".
- Honest: say when a job is not worth doing yourself. Not: every job sounds easy.
- Calm: explain risks plainly. Not: warnings in every paragraph or fear words like "disaster".
Anthropic also recommends saying why a rule exists, because "Claude can generalize from the explanation", and telling the model what to do instead of only what not to do (Anthropic: prompting best practices (external link)). So "write in short paragraphs, because most readers are on a phone" works better than "no long paragraphs".
Form of address
In English, decide whether you speak to "you" directly and whether the company is "we". In German this choice matters even more: the formal Sie or the informal du. Models switch between them in the middle of a text if you do not fix it, so state it once and without exceptions, including in headings and calls to action.
Words to use and words to avoid
Two short lists do more than a long description. The "use" list holds the words your customers use ("lawn" not "turf", "quote" not "estimate"). The "avoid" list holds stock phrases that make text sound generic, words that are wrong for your trade, and competitor names you do not want to mention. Give a replacement where you can, so the model knows where to go.
Typical candidates for the avoid list: "in today's fast-paced world", "let's dive in", "game changer", "unlock", "elevate", "seamless", and the pattern "it's not just X, it's Y". Do not overdo it: forum users warn that a list of fifty bans makes the text stiff and bland. Ten to fifteen entries are enough to start.
Sentence length and formatting
Give rules of thumb, not exact counts: "most sentences under 20 words, one idea per sentence, paragraphs of two to four sentences". Then say how a post is built: an answer in the first paragraph, H2 headings as real questions, lists only for steps or items that belong together, bold only for the one term a reader must not miss. Anthropic notes that the style of your prompt can influence the style of the output, so a guide written in plain paragraphs gets fewer bullet points back than a guide full of them (Anthropic: prompting best practices (external link)).
Facts and claims
This part protects you, and most voice guides forget it. State what the AI must not invent: prices, numbers, customer stories, quotes, guarantees, legal or health statements. Tell it what to do instead, for example "write [price to check] where a price belongs". The person who reviews the post then fills the gaps from real data. Our AI content review checklist shows how to check those points before publishing.
Example paragraphs
Examples carry more voice than any rule. Anthropic calls them "one of the most reliable ways to steer Claude's output format, tone, and structure" and recommends examples that are relevant, varied and clearly marked as examples (Anthropic: prompting best practices (external link)). OpenAI advises showing "a diverse range of possible inputs with the desired outputs" (OpenAI: prompt engineering (external link)).
Use two or three paragraphs you wrote yourself and are happy with, ideally from different content types (a how-to, a news post, a short FAQ answer). Do not use one example only: the model tends to copy its topic and structure instead of its voice.
How to test and refine your voice guide
- Pick a typical topic. Not your easiest one: take a post with facts, a recommendation and a call to action.
- Generate one draft with the guide. Use the same tool and settings you will use later.
- Mark every deviation. Wrong form of address, words from the avoid list, sentences that are too long, invented facts, a tone that is off. Note the line, not just "sounds wrong".
- Fix the guide, not the post. Each deviation becomes a rule, a list entry or a better example. If the same problem appears twice, the rule is unclear: rewrite it.
- Test with a second, different topic. A guide that only works for one topic was tuned to that topic.
- Stop when two drafts in a row need only small edits. Then use it for real posts.
- Check for drift once a month. Compare a current post with the example paragraphs. When something has crept in (a new pet phrase, longer sentences), add it to the guide.
Anthropic suggests a simple check for any instruction: show it to a colleague with little context and ask them to follow it. "If they'd be confused, Claude will be too" (Anthropic: prompting best practices (external link)). For a voice guide, ask a colleague to rewrite one paragraph using only the guide.
How agencies keep several client voices apart
The forum question "can I set up multiple voices?" comes up again and again. The practical answer is strict separation:
- One guide per client, same structure. Every guide uses the same headings (the table above), so nobody has to search, and differences between clients become visible at a glance.
- A short "not this client" note. If two clients are close (two trades businesses, two law firms), note what separates them: "Client A says du, client B says Sie", "client A never mentions prices".
- Separate workspaces. Keep each client in its own project or workspace in your AI tool so instructions and examples do not mix. In Claude, for example, projects are "self-contained workspaces with their own chat histories and knowledge bases", and you can "define project instructions for each project" (Claude help: what are projects? (external link)).
- The client approves the guide. Send the guide with one sample post and ask for approval. It is easier to agree on a page of rules than to argue about every post later.
- Version and date the guide. When a client changes something, update the guide first, then the next post.
More on running client blogs in one place on our page for agencies.
Example: a complete voice guide to copy
The company below is fictional. Replace every detail with your own.
Brand voice guide - Fernhill Gardens (fictional example)
Version 1.0, [date]
1. Who we are
A landscaping firm with eight staff near Bristol. We design, plant and
maintain private gardens. We have done this for many years and say so
without bragging.
2. Audience
Homeowners in and around Bristol with gardens of about 100-400 m2.
They do small jobs themselves (mowing, pruning) and hire us for
design, paving, planting plans and drainage. They are not gardeners:
explain plant names and technical terms in a few words.
3. Voice and tone
- Practical: give the step, the tool, the season and roughly how long
it takes. Not: general advice like "plan ahead".
- Honest: say when a job is better left to a professional and why.
Not: everything sounds easy.
- Calm: explain risks plainly, once. Not: warnings in every paragraph.
- Warm but brief: one friendly sentence is enough. Not: small talk.
Tone by content type: how-to = calm and step by step; seasonal tips =
a bit lighter; prices and contracts = plain and exact.
4. Form of address
Speak to the reader as "you". The company is "we". No first names,
no "folks", no "guys".
5. Words to use
lawn (not turf), quote (not estimate), garden (not outdoor space),
planting plan, drainage, paving.
6. Words and phrases to avoid
"in today's world", "let's dive in", "game changer", "elevate",
"unlock", "seamless", "dream garden", "it's not just X, it's Y",
rhetorical questions as openers, exclamation marks.
7. Sentences and paragraphs
Most sentences under 20 words. One idea per sentence. Paragraphs of
2-4 sentences, because most readers are on a phone.
8. Formatting
Start with the answer in 2-3 sentences. H2 headings as real
questions. Numbered lists only for steps, bullet lists only for items
that belong together. Bold at most once per section. End with one
call to action: "Ask us for a quote".
9. Facts and claims
Never invent prices, numbers, customer stories, quotes or guarantees.
Where a price or figure belongs, write [check: price] and leave it for
the editor. No claims about planning rules or tree protection: write
[check: local rules] instead. Name the season, not a fixed date.
10. Examples of our voice
<example>
Lay a new lawn in early autumn or in spring, when the soil is moist
and not too cold. A new lawn from rolls needs watering every day for the first two
weeks. If you go on holiday in that time, ask a neighbour or wait.
</example>
<example>
Paving a patio yourself is possible, but the work underneath decides
how long it lasts. Without a proper base and a slight fall for
rainwater, slabs sink within a few winters. If you are unsure, ask us
to check the base before you start.
</example>
The example tags follow Anthropic's advice to mark examples clearly so the model can tell them apart from instructions. Keep the whole guide short enough to read in five minutes: if nobody on your team reads it, the AI will be the only one who follows it.
Using your voice guide in Blog Automation
In the AI blog writer of Blog Automation, you store the brand voice once for your company: tone, target group, language (German or English), form of address (Sie, Du or none), default length (short about 600, medium about 1,100 or long about 1,800 words), call to action, words to avoid and a sample text in your style of up to 4,000 characters. The guide above maps onto these fields; the example paragraphs go into the sample text. The brand voice currently applies to the whole account and therefore to every website. For agencies, there is no separate voice per client website yet: differences per client go into the notes for the AI when you start a post. AI posts go to review first by default, so a person checks the voice before anything is published.
Frequently asked questions
Why does AI content sound generic, and how do I make it sound like us?
Because without context, the model writes the average of what it has seen. Give it a concrete audience, do and don't pairs instead of adjectives, a words-to-avoid list and two or three real paragraphs in your style. Then test on a sample post and fix the guide wherever the draft goes wrong.
How long should a brand voice guide for AI be?
Long enough to answer the questions in the table above, short enough that a colleague reads it in five minutes. One to two pages is a good target. Long lists of bans tend to make text stiff, and very long examples get copied instead of imitated.
How do I stop the voice from drifting over time?
Use the same saved guide for every post instead of retyping it, compare a current post with your example paragraphs once a month, and add every new pet phrase or habit to the guide. When the guide changes, give it a new version number.
Can I use one voice guide in ChatGPT, Claude and an AI writer?
Yes, if it is plain text with clear headings. Both OpenAI and Anthropic describe the same basics in their documentation: explicit instructions, context and examples. Check the form of address and the avoid list after switching tools, because each model has its own habits.
Should the AI write my voice guide?
It can help you draft it, for example by describing the style of three of your best posts. But the decisions (audience, form of address, what you never say, which claims are off limits) are yours. Review the draft guide as carefully as a post, and test it before you rely on it.