How to write brand voice guidelines for AI that a model can actually follow

Why generic brand voice guides fail with AI

Brand voice guidelines for AI fail when they are written as adjectives, because a language model acts on facts it can check and patterns it can copy, not on a mood. "Friendly but professional" admits almost any sentence, so a model given only that line writes the average of everything it has read. Google's prompt design strategies page, updated September 17, 2026, says to include the instructions and information the model needs rather than relying on it to already have them, and recommends always including few-shot examples, with the note that prompts without them are likely to be less effective. A guide written for people explains why the brand sounds the way it does. A guide written for a model states what the brand does, in sentences a reviewer can mark right or wrong.

The second failure is that a voice set once fades. Li et al. (2024) had pairs of instructed chatbots talk to each other and report significant instruction drift within eight rounds of conversation, on models including LLaMA2-chat-70B and GPT-3.5, and trace part of it to attention decaying over a long exchange. OpenAI's prompt engineering guide makes the same point from the API side: its instructions parameter applies to the current request only, and when a conversation is carried forward by reference, earlier instructions are not in the context. A guide that is pasted at the start of a chat and never again has, on that evidence, a working life of a few turns.

Those two facts set the design. A brand voice guide for AI is short enough to send with every request, written as checkable sentences, and carries a handful of real examples. The sections below give the fields, the number of examples the vendor pages and the research support, the banned-words rule, how to keep several clients apart, and a test that checks the output rather than the document.

What to include: a field-by-field template

A brand voice guide for AI has seven fields, each written as one plain sentence or a short list, and the table gives the field, how to write it and an example line. The fields follow the six kinds of fact overads keeps in its brand memory (what we sell, who we sell to, how we sound, rules to follow, what worked, and other context), plus the examples. Every line should pass one check: could a reviewer read a draft and say whether this line was followed?

Order matters, and the vendor pages agree on it. Google's page says to put essential behavioural constraints, the persona and the output format in the system instruction or at the very beginning of the prompt, and, when the prompt carries a lot of context, to supply the context first and the specific instruction at the end. OpenAI's guide describes a developer message in four sections, usually in this order: identity, instructions, examples, then context, with context near the end because it changes from request to request. Anthropic's prompting best practices page says to place long documents and inputs near the top, above the query, instructions and examples. For a voice guide the practical reading is: identity and rules first, examples next, and the day's brief or product sheet either last when it is short or first when it is long.

Per-channel limits belong in the guide as well, because a post that fits one network is refused on another. Our table of social media character limits has the current figures, and the guide to cross-posting without it looking copy-pasted covers what changes per network beyond length.

How many examples an AI needs

An AI needs three to five example posts in the prompt to copy a voice, and twenty or more posts to derive the rules from, on the vendor pages and the research read for this guide. The three vendor pages put a number or a range on it; the two papers explain where the number comes from and where it stops paying.

Brown et al. (2020) is where the practice comes from: the paper tested a 175-billion- parameter model given tasks and demonstrations purely as text, with no fine-tuning, and found it performed strongly on tasks it had never been trained for. Agarwal et al. (2024) took the same idea to hundreds and thousands of examples, which they call the many-shot regime, and report significant gains over few-shot across generative and discriminative tasks, including cases where the examples overrode what the model had learned in pretraining and cases comparable to fine-tuning. The cost line in the same paper is the practical limit: inference cost grows linearly with the examples, and a voice guide rides every request.

So the working rule has two numbers. The prompt carries three to five examples in one identical format, which is what all three vendors ask for and what Google warns not to exceed by much. The rules in the guide are derived from a larger set, twenty posts or more, read once to find the patterns. overads draws the same line: its memory reads the newest 20 posts a workspace published, or posts pasted in, and keeps up to 20 one-sentence facts from them, with the instruction to describe the pattern (sentence length, tone, humour, emoji and hashtag habits, how posts open and close, recurring themes, words it likes and avoids) and never to quote a post back as a fact.

Words to use and words to ban

A banned-words list is the most checkable part of a brand voice guide for AI, because a reviewer or a short script can test every draft against it in seconds. Google's page frames constraints as telling the model what to do and what not to do, and the list works best when it keeps to that shape: each ban paired with the word to use instead, so the model has a positive target and not only a hole to avoid.

Four habits keep the list useful. Keep it short, ten to twenty items, because every line rides every request and a long list crowds out the examples. Separate hard bans, such as claims the brand cannot legally make and a competitor's name, from preferences such as a word the founder dislikes, so a reviewer knows which kind of miss matters. Include patterns as well as words: no exclamation marks, no rhetorical question as an opener, no hashtag on LinkedIn. And write the reason next to any ban a new writer would question, because the reason is what lets the model generalise to words you did not list.

A list that bans a word without offering one tends to produce the nearest synonym, which is usually banned in spirit. "Never game-changing" alone yields "transformative"; "never game-changing, say what changed" yields a sentence with a fact in it.

Keeping each client's voice separate

Each client's voice stays separate when each brand has its own guide, its own examples and its own place in the tool, and nothing from one brand rides into another brand's prompt. The reason is in the mechanics above: Google's page says few-shot examples regulate formatting, phrasing and patterning, so an example from client A in a prompt for client B steers B's draft toward A. A shared house guide with per-client exceptions fails the same way, because the exceptions are the part the model drops first.

For an agency with several clients, the rule is one document per client, named for the client, with that client's examples and nobody else's, and one place per client to keep it. In overads that place is the workspace: each client is its own workspace, brand memory belongs to the workspace, and the block sent with a draft is built from that workspace's facts only. An agency can create a managed workspace per client over the API and let the client connect its own channels on a hosted page. A team member's own preferences, such as how they like to be addressed, are kept as person facts apart from the brand facts and ride only that member's turns, so one person's habits never become the brand's voice.

Long instructions get the same treatment. A client whose rules run to pages should not have them in the short guide at all; they belong in a named instruction set the model reads only when it applies. overads calls these skills: each has a name of up to 60 characters, one line of up to 200 characters on when to use it, and up to 6,000 characters of instructions, with up to 40 per workspace. Only the names and when-to-use lines ride every turn, in a block capped at 1,800 characters, and a skill's instructions are fetched when the assistant decides it applies.

Testing that the output is on voice

The test of a brand voice guide for AI is the output, checked the same way each time: a blind sort, a banned-word lint and a length check, repeated after a long chat and after every model change. None of it needs a tool, and it fits in an hour.

1. Take twenty recent posts the team would publish again. Use fifteen to derive the rules and pick the three to five examples from them. Hold back five, with the briefs that produced them.
2. Give the model the guide and the five held-back briefs, and generate five posts.
3. Shuffle the five real posts with the five generated ones and ask a colleague who did not write the guide to sort them. Brown et al. ran the same shape of test with human evaluators and model-written news articles. If the colleague sorts nine of ten correctly, ask what gave each one away and write that as a new line in the guide.
4. Lint the five drafts: banned words, sentence length, emoji count, and the length limit of the channel each was written for.
5. Re-run step 2 at the end of a chat of eight turns or more, with the guide sent only at the start, and compare. Li et al. is the reason: drift showed within eight rounds. If the second set is worse, the guide has to be re-sent with every request, which is how overads does it.
6. Re-run the whole test after a model update. OpenAI's guide recommends pinning production to a specific model snapshot and keeping an evaluation suite so a change in model version is measured rather than noticed.
7. Keep the reviewer's edits. An edit made to three drafts in a row is a missing rule, and it is written as one sentence like the rest.

Approval is where the test runs for real. Our guide to human in the loop AI for marketing covers the approval step itself; in overads every workflow starts out asking before anything goes out, and a post marked for approval waits as pending until a member approves it in the app, as the scheduling and approvals guide explains. The edits made at that step are the raw material for the last item in the list.

Where overads fits

overads is AI marketing automation built only for marketing, and it holds a brand's voice as memory that every writing step reads, rather than as a document someone pastes into a prompt. The memory lives under Settings as What Gloofy knows: every fact is one plain sentence of up to 500 characters, of one of six kinds (what we sell, who we sell to, how we sound, rules to follow, what worked, other), visible, editable and deletable by the team, and a deleted fact is gone from the very next prompt because the block is built fresh each time. Facts can be typed in, or learned in one go from the brand's website (up to 24,000 characters of page text), from the newest 20 posts published through overads or posts pasted in, from a PDF, image or text file of up to 15 MB, or from what another AI assistant wrote when asked to summarise the business. A teach-in keeps up to 20 facts and never the material; a quiet pass after a chat or a workflow run keeps at most 5. The store is capped at 200 facts, the lowest-confidence unpinned ones leave first, and a fact the team added is pinned and never evicted.

The block a model sees is ordered voice first. A chat turn carries up to 1,500 characters of facts; a workflow's writing step, whose whole job is to sound like the brand, carries up to 4,000, in the order how we sound, rules, product, audience, context, what worked. Publish's caption generator reads the same memory before it drafts. Creative Studio keeps a separate brand profile for images and ad copy: it reads a homepage and records the brand name, tagline, three colours, voice and tone, product and audience, one profile per workspace, and a logo uploaded once is used verbatim in every generation. The overads product facts list what each product does today.

What overads does not do:

• It does not fine-tune a model on your posts. The voice lives in the prompt, and overads' text AI runs on Gemini, not on GPT or Claude.
• It cannot read your post history from X or LinkedIn, because neither platform offers that to an app like this. Paste the posts, or let it read what you published through overads.
• It does not score a draft for voice. A person approves it, and the blind sort above is run by hand.
• It keeps at most 200 facts and sends at most 4,000 characters of them with a draft. A thirty-page brand book has to be distilled; long instructions go into a skill of up to 6,000 characters that is read only when it applies.
• Publish is live on X, LinkedIn, Bluesky, Mastodon and Pinterest only. It cannot post to Facebook or Instagram today, and it holds no write access to any ad account.

A team with thousands of labelled posts and an evaluation suite that wants a fine-tuned model is better served by a model vendor's tuning service. A team whose guide is mostly visual, type and layout rather than words, needs a design system rather than a voice guide. A solo writer who drafts everything by hand needs no guide for a model at all. overads fits a team or an agency that wants drafts written in the brand's voice on a schedule, from workflow templates that ask for approval before anything goes out. The API and the MCP server are included on every plan, and plans are on the pricing page.