How to write brand guidelines AI tools will actually follow

Brand guidelines were designed for humans. They live in a PDF, they explain things visually, they assume the reader can interpret “use a warm, confident tone” and know what that means. That worked fine when the only people reading your brand guide were designers, copywriters, and the occasional intern. Now, there is a third category of reader that your guidelines were never written for: AI tools. If you want Choco Media-calibre AI brand guidelines that actually shape how your AI tools write and generate, the document needs to change — not the tools.

We work with marketing teams and founders who are already using AI for content production, social copy, email drafts, and increasingly image generation. The consistent complaint is the same: the AI output does not sound like us. Sometimes it is close. Often it is generic. And the reason is almost never the model — it is the instructions. Vague, human-readable brand documentation does not translate directly into AI-usable constraints. This post is about how to change that.

What follows is the set of changes we make to brand guidelines when a client is actively briefing AI tools. These are not theoretical — they are the fields, formats, and examples we add to make brand guidelines machine-readable without making them unreadable for humans.

Why standard brand guidelines fail with AI tools

Most brand guidelines describe the brand in terms of feeling and intention. “Bold but approachable.” “Warm without being informal.” “Authoritative but never stiff.” These descriptions make sense when you read them alongside visual examples, know the company, and have professional judgement. An AI tool has none of that context by default.

When you paste a section of a typical brand guide into a chat model and ask it to write in that voice, the result is usually acceptable. The model is good at general tasks. But “acceptable” is not “on-brand,” and the gap compounds at scale. If you are producing 30 pieces a month with AI assistance, a 10% drift from your actual voice becomes a meaningful consistency problem.

• Feeling descriptions are interpreted differently each session. “Conversational but intelligent” means something different every time a model processes it, because there is no anchoring example to define the range.
• Visual brand elements have no text equivalent in your guidelines. Your typography choice and colour palette do not tell an AI what sentence structure you prefer or how you handle technical jargon.
• Prohibitions are often absent or vague. “Avoid corporate jargon” does not tell an AI which specific words and constructions you object to.
• There is no example library. A human reading good and bad examples calibrates instantly. AI benefits from exactly the same input — it just needs it provided explicitly.

The fix is not to throw out your existing guidelines. It is to add a section specifically designed for AI tools — a machine-readable layer that sits alongside the human-readable one.

The voice metadata block: the most important addition

The single highest-leverage change is writing an explicit voice metadata block at the top of any prompt or brief that involves AI-generated copy. Think of it as a compressed version of your brand voice that a model can parse reliably.

What the metadata block contains

• Tone descriptors with anchoring examples: Instead of “warm,” write “warm — like a knowledgeable colleague explaining something, not a customer service script.” The analogy grounds the instruction.
• Person and register: “First-person plural (we/our), direct address to the reader (you), no passive voice except when describing processes.”
• Sentence structure preference: “Medium-length sentences. Avoid compound sentences joined by more than one and/but. Paragraphs of 2–4 sentences.”
• A banned words list: This is not optional. Write out the specific words and phrases you never use: “game-changer”, “unlock”, “leverage” as a verb, “in the realm of”, “journey” in a business context, “delve”.
• A positive vocabulary list: The words you actually use. Industry terms you prefer. Phrases that are distinctively yours.

We typically format this block as a short numbered list at the top of any AI brief, not buried in a PDF appendix. If it is not in front of the model at the moment of generation, it does not reliably influence the output.

The brand guidelines document is for understanding. The voice metadata block is for production. They serve different purposes and should be treated that way.

Worked example: converting a voice description into AI-usable instructions

Here is a typical brand voice description we encounter: “Our tone is professional yet approachable, informed by our deep expertise but never condescending. We communicate with warmth and clarity.”

That is pleasant to read. It communicates nothing specific to an AI. Here is the same intention rewritten as machine-usable instructions:

• Write in first-person plural. We do things. We see things. Not “one might consider” or “it can be argued.”
• Sentence length: 12–22 words on average. Shorter for emphasis. Never more than 35 words in a single sentence.
• No jargon without explanation. If you use a technical term, define it in the same or following sentence.
• Never start a section by restating the heading as a sentence. Begin with the substance.
• Avoid: “it’s important to note”, “needless to say”, “at the end of the day”, “moving forward”, “in today’s landscape”.
• Preferred openers: direct observations, specific data points, short questions, counter-intuitive statements.

These are instructions an AI can follow consistently. Test them with a short sample — if the output is materially different from your actual brand voice, add more examples. Iteration here is normal. The goal is a block you can paste into any AI brief and get reliably on-brand output within one or two revision passes.

Our guide to building a brand voice document AI can actually follow covers the full document structure — the AI metadata block described here is the production-facing layer on top of that foundation.

Example sets: the fastest way to calibrate an AI

If you add one thing to your brand guidelines for AI use, make it an example library. Two or three annotated examples of on-brand writing will do more than five paragraphs of voice description.

How to structure examples for AI briefing

• Positive examples first: Two excerpts (a paragraph each) from your best existing content. Label them “On-brand example 1” and “On-brand example 2.” No annotation needed — the model learns from the sample itself.
• Negative examples second: One or two examples of writing that misses the mark, with brief notes explaining what is wrong. “Too formal — reads like a press release” or “Too casual — uses slang we don’t use” is enough.
• Transformation examples (optional but powerful): Take a generic sentence and rewrite it in your brand voice. Show the before and after. These are particularly effective for training models on register.

The example library does not need to be long. Four to six examples — three positive, two negative, one transformation — is typically sufficient to anchor a model’s output reliably. Keep the examples in plain text format so they can be pasted directly into prompts without stripping formatting.

Making guidelines work for AI image generators

Text-based brand guidelines rarely address image generation because image generation did not exist when most were written. If your team is using Midjourney, DALL-E, or Firefly for marketing visuals, your guidelines need an additional section.

What to include for image generation

• Visual style descriptors in prompt-compatible language: “Photorealistic, natural lighting, warm tones, no lens flare, no stock photo feel” is immediately usable. “Clean and modern” is not.
• Colour references in words: Describe the palette in natural language that models understand: “muted earth tones, sage green, warm off-white, avoid saturated primaries.”
• Subject and composition defaults: What kind of people appear in your visuals? What settings? What is always absent? (“No skylines, no boardrooms, no handshake photos.”)
• Negative prompt library: The most underused tool in AI image generation. Write a standard negative prompt for your brand: the visual elements, lighting styles, and aesthetic directions you consistently want excluded.
• Aspect ratio and format standards: What proportions do you use? 16:9 for hero images? 4:5 for Instagram? Include these so any team member briefing image generation tools defaults to the right format.

We help clients build this section as part of our brand identity and guidelines work. The image generation layer is now a standard deliverable in any brand identity project we run.

The governance layer: keeping guidelines current as AI tools evolve

AI tools change faster than brand guidelines are typically updated. A prompt pattern that works reliably with one model version may produce noticeably different output after an update. This is not a reason to avoid AI tools — it is a reason to build a lightweight governance layer into your brand documentation process.

• Assign an owner: One person is responsible for keeping the AI-usable brand layer current. In a small team this is usually whoever runs content production. In a larger team it may be the brand manager.
• Set a review cadence: Quarterly is appropriate for most organisations. Test the voice metadata block and example library against a new piece of content and update if the output has drifted.
• Version the document: Use dates, not version numbers. “Brand voice metadata — updated 2026-Q1” is clearer than “v2.3.” If the guidelines are in Notion or a shared doc, the edit history handles this automatically.
• Capture drift as it happens: When a team member notices that AI output does not match the brand voice, that is a signal. Create a process for those observations to reach the brand guidelines owner — a Slack channel, a Notion comment, a monthly note in a shared doc.

For teams running AI at scale, connecting brand documentation to the production workflow directly — rather than relying on individuals to remember to include the voice block — is the next step. Our AI automation services cover this as part of a broader content operations build.

Common mistakes when adapting brand guidelines for AI

We have seen the same mistakes repeated across client engagements. They are worth naming directly.

• Writing the AI layer once and never revisiting it. Model behaviour changes. Guidelines need maintenance.
• Making the voice block too long to include in a prompt. If the metadata section is 800 words, nobody will paste it consistently. Aim for 150–250 words for the production-facing version. Keep the full version separately.
• Relying on the model to infer from the company name alone. “Write in the voice of [Brand X]” only works if the model has seen substantial public content from that brand. For most businesses, it has not.
• Using only positive constraints. Telling an AI what you do want is half the job. The banned words and negative examples are what actually narrow the range of acceptable output.
• Not testing before deploying. Run the metadata block against five sample prompts before declaring the guidelines ready. Adjust based on what you see, not what you assumed would work.

A practical starting point: the minimum viable AI brand block

If you are starting from scratch — or adapting existing guidelines with limited time — here is the minimum viable AI brand block. Fill in the brackets with your own content and test it immediately.

• Voice in one sentence: [Describe the tone as a human analogy: “like a [role] explaining something to a [audience], not a [contrast].”]
• Person and register: [First/second/third person. Formal/informal. Active/passive preference.]
• Sentence structure: [Average length target. Paragraph length. Any structural constraints.]
• Banned words and phrases: [List 8-12 specific words and constructions you never use.]
• Preferred vocabulary: [List 5-8 words or phrases that are distinctively yours.]
• Two positive examples: [Paste two real excerpts from your best content.]
• One negative example: [Paste one example of the kind of writing you want to avoid, with a one-line note on what is wrong.]

That block, consistently applied, will produce a noticeable improvement in AI output quality within the first week. Refine from there as you gather examples of what is still not right.

If you are ready to build a brand guidelines system that works reliably for your AI production workflow, get in touch — it is one of the most practical things a brand can do right now, and the setup is faster than most teams expect.