Brand guidelines for AI: turn the brand book into decisions

A brand book is written for designers to interpret. It says “sound human” or “confident, never arrogant”, and a good designer turns lines like that into dozens of small decisions without writing any of them down. An assistant given the same page has nothing to interpret it against, so it guesses, and the guess changes from one chat to the next and from one colleague to the next. That is why uploading the PDF and hoping produces copy that is nearly on brand and never quite the same twice.

What an assistant can follow is a decision: the company name spelled one way, a colour as a code, a line you have actually published that shows the tone, a word you never use. Anthropic’s advice for instructions is to make them concrete enough to check, a specific indentation width rather than “format code properly” [1], and to explain the reason behind a constraint, because Claude generalises from the explanation [2]. Brand rules work the same way. “Sound human” cannot be checked, while “write ‘we’, never ‘the company’” can.

The second gap sits between the book and what the company actually publishes. Websites, proposals and posts drift away from a brand book in ways nobody records, and some of that drift is the company learning what works. Those differences are where the real rules live, and an assistant given only the book cannot see them.

Open a new chat in the work account your company has approved, not a personal one. Attach the brand book and six things you have published that carry the brand in words, such as a web page, a proposal, a post, a press release and a customer email. Pick pieces that differ from each other: Anthropic’s prompting guidance counts relevant, varied examples among the most reliable ways to steer format, tone and structure [2]. Then type this.

The first half asks for the book rewritten as rules a stranger could apply. “Without us” is the test for every line: if someone outside the company could not follow it, it is still a description. The four items name what the rules must contain, so the answer comes back as decisions rather than a summary of the book’s mood.

The second half is the part that earns its place. Asking where the book disagrees with what you published turns the assistant from a reader into an auditor, and “ask me which is right” stops it settling the question by itself.

Answer its questions in the same chat. When the list is settled, ask it to save the rules as a skill. Claude and ChatGPT can both create a skill from the conversation you have just had [3, 4], and the skill is where tomorrow’s draft starts.

Each phrase asks for a different kind of decision, and each one closes a gap that a brand book tends to leave open.

The disagreement list is the most useful thing the prompt produces, and it can be uncomfortable to read. Typical entries are the company name written two ways across the six pieces, a word the book bans sitting in a website headline, a tone rule the proposals ignore and a colour on the site that the book never lists.

Each entry needs a decision from whoever owns the brand, and the decision goes into the rules with its reason. Sometimes the book wins and the published piece was a mistake. Sometimes the published work wins, because the company learned something after the book was signed off and the book was what had gone out of date. Either way the question is answered once, and every draft after that starts from the answer.

If the answers come from more than one person, record who decided each one. The next disagreement then goes straight to the right person instead of round the whole marketing team.

The rules work in both directions. Once the skill exists, the list that shapes a new draft can also check one somebody else wrote, including copy that arrives from an agency. Ask for quotes, so you see each breach rather than taking a verdict on trust.

The last instruction matters most. A case the rules do not cover is a question for the brand owner, and the answer becomes the next line in the skill. OpenAI’s guidance on reusable skills makes the same point in its own terms: when a skill gets something wrong, ask for the correction to be added to the skill [7].

Work the brand team has already rejected is a good source of rules, because every rejection came with a reason. Feed it in once the first version of the skill exists.

A contradiction here is one more unwritten rule, found late. Settle it the same way, once, with the owner. Anthropic separates skills that add a capability the model lacks, which may stop mattering as models improve, from skills that encode how a team works, which last but have to stay faithful to how the team really works [8]. A brand skill is the second kind, so it needs feeding as the brand moves.

Both products use the same structure. A skill is a folder whose one required file, SKILL.md, carries a name and a description followed by the instructions [9, 10], and the assistant reads each skill’s name and description first, loading the rest only when it decides the task matches [10, 11]. Creating, switching on and sharing one differ by product and plan.

The Claude column is from Anthropic’s help centre pages on using skills [12] and on provisioning them [13], its skills overview [11], its tutorial on creating a skill through conversation [3] and its article on uploading files [14]. The ChatGPT column is from OpenAI’s Help Center pages on skills [4] and on uploading files [15], its guide to building skills [10], OpenAI Academy [6] and its docs pricing table [16]. Menu labels move between releases, so check each step in your own account.

If the brand book is a designed PDF, the last row matters more than it looks. A colour swatch with its code printed in the picture, or a logo page with its spacing note set as an image, may not reach the assistant at all. Paste those rules in as text, or check the rules it returns against the book page by page before you save them.

A brand skill in one person’s account helps one person. The point is every assistant in the company starting from the same rules, and both vendors give administrators a way to get it there.

In Claude, a custom skill is invisible to colleagues until you share it. Shared skills are view-only, stay off until each recipient turns them on, and pick up your updates automatically [12], so a rule the brand owner changes reaches everyone who has the skill without a memo. On Team and Enterprise, owners can provision a skill for everyone from Organization settings > Plugins & skills, where it is on by default and each person can switch it off; giving it to only some people is Enterprise only, through a plugin assigned to a group [13]. In a ChatGPT workspace you can publish a skill to the workspace library, and admins can see each skill’s owner, access, users and invocations over the last 30 days [4].

Before a skill goes company-wide, OpenAI’s admin guidance asks for its source, an accountable owner, its intended audience and a review date, a test with non-sensitive data, and a record of who owns re-review and retirement [17]. Anthropic’s course on sharing what you build asks the same of shared plugins: one named owner, a test run before every publish and a regular review [18]. For a brand skill the owner is whoever already owns the brand book.

Which assistant the company standardises on is a separate decision, and our side-by-side of what ChatGPT Enterprise and Claude Enterprise publish on price, admin controls and data location covers it.

None of these is a reason to skip the exercise. Each is a reason to keep a person reading what goes out.

Rolling out assistants and the skills that sit on them is core 1AYM work. We roll out ChatGPT Enterprise and Claude Enterprise, and at a large international marketing agency we hold the lead architect role on its AI platform and contribute engineering to its internal Claude Code skill platform.

For a brand skill we would run the exercise on this page with your brand owner, then set the skill up as a shared asset: one owner, published or provisioned to the people who write in the company’s name, a test set built from real drafts and a review date. That is the rollout side of our AI governance implementation work, and it can be a small fixed-scope statement of work, which can start within a day of the scope being signed. If that would help, book a call from the end of this page or email enquiries@1aym.com.