Turning Brand Guidelines Into Rules AI Tools Can Follow

Most companies that now use AI for posts, emails, product descriptions and proposals already own a brand book. It explains the logo, the colours, the fonts and how much clear space to leave around the mark. Then a sales coordinator asks an AI tool for a LinkedIn post, a marketing executive asks another tool for an Arabic caption, and the agency uses a third. Three different voices come back, each one fluent, none of them yours.

The fix is not a better prompt typed from memory each time. It is a short, written rule set for language that people and AI tools can both follow, kept in one place and loaded into every tool your team uses. This guide shows what that sheet should contain, how to phrase the rules so a model can apply them, and how to keep it working across staff and agencies.

Why the brand book does not keep AI copy on brand

Visual guidelines are written for designers, who share a trained eye and can read “confident, premium, modern” and produce something reasonable. A language model reads the same adjectives and fills the gaps with its own defaults: generic enthusiasm, long introductions, stacked superlatives and the same handful of stock phrases every other company is publishing.

Three gaps cause most off-brand AI copy:

• No terminology. The tool does not know, for example, that you say “fire suppression system” and never “fire extinguishing machine”, or that your Arabic product name is fixed and must not be translated.
• No limits on claims. Without explicit rules, models can easily produce “the best in the region”, “fully certified” or “guaranteed results”, which can expose you commercially and, in some sectors, legally.
• No channel or language rules. A tone that works in a tender cover letter is wrong for an Instagram caption, and the Arabic version needs its own rules rather than a literal translation of the English ones.

If your visual guidelines are also incomplete, start there; our guide to what brand guidelines must contain, including the Arabic rules covers the design side. This article covers the language system that sits next to it.

The six parts of an AI-ready voice and terminology sheet

We recommend keeping the sheet short enough that people actually read it: in our experience four to eight pages, written as rules and examples rather than essays, is enough. These are the sections we suggest.

1. Voice principles that can be checked

“Professional and friendly” cannot be tested. “Use ‘we’ and ‘you’; sentences under 25 words on average; no exclamation marks in corporate channels” can. For each trait, write one sentence of definition, two do’s, two don’ts and one example sentence. If a reviewer cannot point at a line in a draft and say which rule it breaks, the rule is too vague.

2. Approved terms, in both languages

In practice this is often the section that saves the most editing time. Build a two-column glossary: the English term and its approved Arabic equivalent, plus a third column for terms to avoid. Include:

• Company and product names, with exact spelling and whether they are ever translated or transliterated
• Service names as they appear on your website and quotations
• Technical terms your buyers use, including the English terms Arabic-speaking engineers prefer to keep in English
• Partner and principal brand names you represent, written exactly as the partner requires
• Job titles and how you address people in each language

For distributors and agents, this list often matters more than the voice principles: a principal brand will notice a mistranslated product name long before it notices tone.

3. Banned and controlled claims

Split this into two lists. The banned list is short and absolute: superlatives you cannot prove, guarantees, competitor comparisons, and any phrase your legal or compliance team has ruled out. The controlled list covers claims that are allowed only when they match an approved source: certifications, approvals, years in business, project counts, client names and warranty terms. Give the tool the approved source itself (a fact sheet with the current numbers) and the rule “use only figures from this sheet; if a figure is missing, leave a placeholder in square brackets”.

That placeholder rule is one of the most useful lines in the whole document. It turns an invented number into a visible gap that a person has to fill.

4. Tone by channel

Write one short block per channel: purpose, length, structure, formality and call to action. A typical B2B set covers LinkedIn company posts, executive LinkedIn posts, Instagram and Facebook captions, WhatsApp replies to inquiries, email follow-ups, website service pages and proposal or tender text.

5. Arabic and English style rules

Treat Arabic as its own language with its own rules, not as the output of a translation step. Decide and write down:

• Register. Modern Standard Arabic for corporate copy, and whether any channel (for example, consumer social posts in Egypt or Saudi Arabia) may use a light dialect, with examples of where the line sits.
• Numerals and mixed strings. Turn the digit and mixed-text decisions in your brand guidelines (see the guide linked above) into instructions, for example “Use Western digits in all Arabic social posts”.
• Punctuation and spelling. Arabic comma and question mark, and your choices on common spelling variants, so the same word is not written two ways across posts.
• Direction of work. Whether Arabic copy is written first, written in parallel, or adapted from English. Adapted copy should be rewritten for meaning, not converted line by line. Our guide to where AI falls short in Arabic copy shows what to check.

For English, decide British or American spelling, how you write dates and currencies (SAR, AED, EGP, QAR), and capitalisation of job titles and service names.

6. Sample rewrites

In our experience, examples convey a voice better than adjectives. We suggest five to ten pairs: a typical generic draft and the approved rewrite, ideally in both languages and across several channels. For example, for a hypothetical training academy (an illustration, not a real client):

Generic draft: “We are thrilled to announce our cutting-edge courses that will revolutionise your career!”
Approved rewrite: “Our new project management course starts in Riyadh on [date]. The course outline and fees are available on request.”

Annotate each pair with the rules it applies. Those notes also train your team.

How to phrase rules so a model applies them

How a rule is written affects how consistently it is applied. These are the habits we recommend:

1. Write instructions, not descriptions. “Start with the main point” is easier to apply and check than “our voice is to-the-point“.
2. Say what to do instead. “Do not use ‘solutions’ on its own; name the system or service” gives the tool a replacement, not just a prohibition.
3. Use numbers where you can. Word ranges, hashtag limits and sentence length are easier to follow and to check than “concise”.
4. Set a priority order. State what wins when rules conflict, for example: accuracy and approved claims first, then terminology, then channel format, then voice.
5. Tell the tool what to do when it does not know. Placeholders for missing facts, and a request to list open questions at the end of the draft.
6. Keep one rule per line. Long paragraphs hide rules; short lines are easier for both models and reviewers to apply.

Where to load the rules in the AI tools your team uses

The main assistants let you save instructions, and some let you attach reference files, so you are not pasting the guide into every chat. These features change often; the points below reflect each vendor’s help page as of October 2026, so check the current help page before you roll out.

• ChatGPT custom instructions apply to new chats, and OpenAI’s help page lists two fields with a limit of 1,500 characters each (OpenAI Help Center). That is enough for a summary of your rules, not the full sheet.
• Claude Projects give a workspace its own instructions and knowledge base of files and text; Anthropic’s help centre says sharing projects with colleagues is available on Team and Enterprise plans, with view or edit permissions (Claude Help Center).
• Gemini Gems let you save instructions, add files from your device or Google Drive, and share the Gem with others. Google’s help page also states: “Starting in November 2026, Gems will go away on personal Google accounts and automatically transition to skills.” That change is due to begin within days of publication, so if your team uses personal accounts, check how your saved instructions and files carry over. Google adds that work and school accounts “may be subject to different terms” (Gemini Apps Help).

Whatever the tool, use the same structure:

• One master document, owned by marketing, with a version number and date. This is the only version anyone edits.
• A short instruction block (a few hundred words, cut to fit any character limit) that summarises the non-negotiables: priority order, banned claims, terminology rule, placeholder rule, language rules. This goes into the instruction field of each tool.
• Reference files attached as knowledge where the tool allows: the full sheet, the glossary, the approved fact sheet and the sample rewrites.

Keeping staff and agencies on the same version

The sheet fails quietly when it drifts. One team keeps last year’s version in a shared folder, the agency works from a PDF someone emailed, and a new product name never reaches either. A light governance routine prevents most of this:

• Name one owner who approves changes and publishes new versions.
• Keep a change log at the top of the document: date, what changed, why.
• Update the tools when the document changes. Shared projects or Gems mean one update reaches everyone who uses them; personal instructions have to be updated by each person, so list who holds a copy.
• Give agencies the same pack as internal staff, and ask them to confirm which version they are working from.
• Keep human approval. The sheet reduces editing, it does not replace review. Someone who knows the business still checks facts, claims and Arabic quality before anything goes out, ideally through a written AI content approval workflow. Google’s guidance on generative AI content asks publishers to focus on accuracy, quality and relevance and to fact-check and review AI-generated content before publishing (see also our note on Google’s updated AI content guidance).

At RX, this sits naturally inside brand guidelines work: the visual system and the language system are built together, and clients approve the content rules the same way they approve the logo. For teams that run monthly bilingual content, the same sheet becomes the brief for social media and content production.

Test it before you roll it out

Before the sheet goes to the whole team, we recommend a simple test to show which rules are not landing.

1. Write ten realistic requests your team makes every month: a product post, a WhatsApp reply to a price inquiry, an exhibition invitation, a service-page paragraph, an Arabic caption, a follow-up email after a quotation.
2. Run each request through every tool your team uses, with the instruction block and files loaded.
3. Score each output against five checks: approved terms used, no banned or unsupported claims, correct channel format, correct Arabic or English style rules, and voice consistent with the sample rewrites.
4. For every failure, find the rule that should have prevented it and rewrite it: more specific, a number added, or a sample rewrite that shows it.
5. Repeat until most outputs need only light editing, then publish version 1.0 and schedule a review each quarter or whenever products, services or partners change.

Keep the test requests. When a tool updates its model or features, rerun them to see whether your outputs have shifted.

If you can only draft two sections this week, make them the approved-terms glossary and the banned and controlled claims list. Build both from last month’s published posts and quotations, then run the ten test requests against them before you write the rest.