AI brand guidelines: how to make your brand book work in Claude and ChatGPT

From a PDF to documents with sources, rules that pass down to every sub‑brand, and one approved version that every AI tool on your team reads.

Short answer

AI brand guidelines are brand rules written so that an AI tool can read and apply them: one topic per document, each rule stated with an example, a source and a version. A PDF brand book fails at this because it reaches the AI as a pasted copy that nobody updates. The fix has four parts. Split the book into documents with sources. Mark which rules every sub‑brand must keep. Serve one approved version to every AI tool, for example over MCP. Then decide which parts of a layout AI may fill and which stay locked.

Why don’t AI tools follow a PDF brand book?

Because the PDF reaches the AI as a one-off copy, and the rules inside it were written for a designer who already knows the context.

A brand book is laid out for people. Rules sit in captions next to mood images, logo misuse is shown as a picture, and the tone of voice is a page of adjectives. When someone uploads the file to a chat, the model gets the extracted text with the layout gone. “Confident but approachable” gives it a feeling to aim for and no way to check the result.

The bigger problem is the copies. Each person uploads their own file: last year’s version on one laptop, a slide export on another. When the brand team changes a rule, none of those copies change. The model also fills every gap it finds. If the guidelines never say how long a LinkedIn post should be, the model picks a length, and the team reads the result as “the AI ignored our guidelines”.

What does an AI-ready brand guideline look like?

It is a set of short documents, one per topic, where each rule is a sentence you could check a draft against.

  • One topic per document. Positioning, messaging, tone of voice, each product, each spokesperson and each channel get their own document, so an AI tool can fetch the one it needs.
  • Rules with examples. Write pairs of do and don’t, each with a real sentence. “Warm” means little to a model. “No exclamation marks, one emoji at most” can be checked.
  • A source for every rule. Note the page of the brand book, the workshop or the campaign it came from. When the AI cites a rule, a person can trace it, and when two sources disagree, the brand team sees it first.
  • Limits as numbers. Headline length, number of hashtags and minimum colour contrast belong in the document as values.
  • Version and status. Every document shows whether it is a draft or approved, and when it last changed.
Figure 1. One document per topic, as it looks in Monobase. Lumo is a fictional brand.
  1. The path: which level the rules come from
  2. Status and version
  3. A rule with an example
  4. Limits as numbers
  5. The sources behind the rules

Visual rules work the same way. Colours, type and spacing are easiest for a tool to use as design tokens: named values such as color.action.primary that software reads as data. The W3C Design Tokens Community Group published the first stable version of its format, 2025.10, in October 2025 (W3C), so there is now a shared file format for them.

How should rules pass down to sub‑brands?

Treat the brand as a tree and mark which documents are binding. The group sets the rules every brand below it keeps, and each sub‑brand, product, spokesperson or channel adds only what is its own.

Take Lumo, a fictional drinks group. Lumo Group sets the positioning and a tone of voice. Lumo Drinks, its soda brand, adds the voice for Lumo Citrus and Lumo Berry. The CEO’s LinkedIn adds rules for writing in the first person. When an AI writes for Lumo Citrus, it should read the group’s binding documents and the Lumo Drinks additions, and nothing from Lumo Water.

  • A lower level adds and never overrides. If a sub‑brand needs an exception to a binding rule, change the group document, so everyone can see the exception.
  • Show the path. A heading such as “Lumo Group / Lumo Drinks / Tone of voice” tells the person reading an AI answer which level a rule came from.
  • Don’t copy group rules into sub‑brand documents. Copies drift apart within a few edits.

How do you keep every AI tool on the approved version?

Keep the guidelines in one place that AI tools read while they work, instead of in files people upload. Teams use four methods today, and they differ mainly in who updates the copy the AI reads.

Four ways to give an AI tool your brand guidelines
MethodWhere the rules liveWhen a rule changesWorks best for
Upload the PDF to a chatIn that one chatEveryone uploads the new file againA one-off task
Project or custom GPT with filesFiles attached to the projectThe project owner replaces the filesOne small team with one brand
Skill file (SKILL.md)A folder of instructions the AI loads when a task needs itSomeone edits the file and shares it againRepeated tasks with a fixed method
MCP server in front of the guidelinesIn the source system; the AI fetches documents when it needs themApprove the change once, and every connected tool reads it on its next requestSeveral brands and several AI tools

Anthropic introduced Agent Skills in October 2025: folders of instructions, scripts and resources that Claude loads when a task calls for them (Anthropic). They suit a method that rarely changes, such as how your team structures a press release. MCP, the Model Context Protocol, is an open standard Anthropic introduced in November 2024 for connecting AI assistants to the systems where data lives (Anthropic). Claude supports it, and OpenAI documents how to connect an MCP server to ChatGPT (OpenAI). The two combine well: a skill holds the method, and MCP supplies the current rules.

Whichever server you connect, check four things before your team relies on it:

  • It serves the approved version of each document, so edits in progress reach the AI only once someone approves them.
  • It is read-only for the AI by default, so a chat can’t change a rule unless you allow it.
  • Access follows your roles, so an agency sees only the brands it works on.
  • The AI tool can name the document it used, so a reviewer can check the answer.

Monobase meets all four. Each brand you serve can sit in its own workspace with its own MCP address, so a connection reads only the workspace of the person who signed in.

Figure 2. Claude fetches the approved documents over MCP while it writes and lists them as sources, the group’s binding rules first and then the sub‑brand’s own. A reviewer can open each one. Claude stands in for any MCP client; Lumo is fictional.

In Monobase, for example, the MCP connection gives Claude and ChatGPT read-only access to a workspace’s brand documents and design system by default; an owner can switch on writing for owners and admins. Once a document is approved, AI tools read that version; new edits wait in Monobase until the next approval. A document nobody has approved yet is read as its live draft.

What should you lock, and what can AI fill?

Lock what makes the brand recognisable and what carries legal risk. Let AI fill the parts that change with every format and message, within limits you set, and keep a person on approval.

What to lock and what to let AI fill in a brand template
LockLet AI fill, within limits
Logo, safe zone and layout gridThe headline, up to a set number of characters
Colour roles and type stylesBody copy and captions in the approved voice
Product names, prices and legal claimsBackground images from an approved image generator
Photos of real people and mandatory disclaimersResizing and reflowing the layout for each format

Write these decisions into the template itself, not into another document. The designer’s master file is the right place: each layer is either locked or open, and each open layer carries its limit. An AI that fills such a template can only change what the template leaves open. Monobase builds its brand templates this way: designers mark in a Figma master which layers AI may fill, and agents fill only those.

Locks reduce what can go wrong. They don’t replace review. The AI drafts, and a person approves before anything is published.

Where should a brand team start?

Start with the documents AI tools read most often, and move outwards from there.

  1. Collect what you have: brand books, strategy decks, workshop notes, your website and recent posts.
  2. Pick the documents your team’s AI tools use most. For most marketing teams that is tone of voice, messaging and product facts.
  3. Rewrite them as rules with examples and sources. Where sources disagree, write the question down instead of guessing.
  4. Decide which documents are binding at group level and which belong to a sub‑brand.
  5. Connect your AI tools to one approved copy, and ask people to stop uploading the PDF.
  6. Turn your most-used design master into a template with locked and open layers.

FAQ

Do we have to rewrite our brand book?

No. Most of the work is restructuring: splitting the book by topic, turning adjectives into rules with examples and adding sources. Your positioning and visual identity stay as they are.

Is MCP only for developers?

Setting up a server is a technical job. Using one is not: once it is connected, people ask Claude or ChatGPT as usual, and the tool fetches the brand documents it needs.

Can an AI tool change our guidelines over MCP?

Only if the server allows it. For brand guidelines, read-only access is the safer default, and changes go through the people who own each document.

Should we still keep a PDF?

Keep one for partners who need a file, and export it from the documents, so there is still only one source to update.

Do image rules work the same way?

Yes. Write what is allowed, show examples of do and don’t, and tune your image generator or photo brief to those rules. The template keeps the layout around the image locked.

Sources

  1. Anthropic, Introducing the Model Context Protocol, 25 Nov 2024.
  2. Anthropic, Introducing Agent Skills, 16 Oct 2025.
  3. OpenAI, Add custom MCP server, developer documentation, checked 21 Oct 2026.
  4. W3C Design Tokens Community Group, Design Tokens specification reaches first stable version, 28 Oct 2025.

Written by

Jiří Kadlčík, co-founder of Monobase

Monobase is made by Symbio, an independent design and creative agency in Prague. The Symbio team sets Monobase up on your brand with you.

About Monobase and Symbio
Published
21 Oct 2026
Last checked
7 Nov 2026
Examples
Lumo is a fictional brand

See your own brand book as documents with sources.

A 30-minute demo of the brand guidelines part of Monobase, including how Claude or ChatGPT reads it. Send a brand book with your request and we’ll run part of the demo on it.

More from Insights. Two pieces in this series are being written.

All Insights

Same brief, three set-ups

ComparisonBrand guidelines in AI toolsIn writing

We gave Claude the same brand book three ways

As a PDF in a chat, as a skill file and through an MCP server. The same brief in each, the outputs side by side, and what each set-up does when the brand team changes one rule.

AI-readiness · 2026

ResearchAI-readiness reportIn writing

AI-readiness of public brand guidelines, 2026

Published brand guidelines scored against ten criteria. The method goes up first, then every score, and each brand in the study can send corrections.