HrvatskiPart 1
Who we are
One sentence on what you do and for whom. The agent uses this exact sentence.
Brand OS Portal
Brand guidelines for AI agents: rules any tool can follow.
A chat tool writes a draft. An agent sends it. So agents need rules on what they may do, not just voice.
Brand guidelines for AI agents are a short plain-text file with five parts: who you are, approved facts, voice rules, what the agent may never say, and what it must ask a person about first.
Write it once, in plain words, and load the same file into every agent and tool you use. Then it does not matter which tool runs the job: the brand stays the same.
By Vedran Obradović · Updated 4 November 2026
Why agents need different brand rules.
A chat tool waits for you. An agent works on its own.
| Chat tool | AI agent | |
|---|---|---|
| What it does | Writes a draft for you | Drafts, posts, sends or updates |
| Who reads it first | You | Maybe nobody |
| Biggest risk | A weak draft | A wrong claim in public |
| Rules it needs | Voice and facts | Voice, facts, limits and when to stop |
Tool-specific setup lives on other pages: Claude, ChatGPT and AI-ready brand guidelines. This page is about rules that work in any agent.
The five parts of brand rules for agents.
Part 2
Approved facts
Numbers, names and claims with a source. The agent may use only these.
Part 3
Voice rules
Three to five rules, each with a right and a wrong line.
Part 4
Never say
Banned words, banned claims, competitors it may not mention, promises nobody may make.
Part 5
Stop and ask
What the agent must hand to a person: new claims, prices, complaints, legal topics, anything public for the first time.
Free tool
Can AI tools read your brand today?
Type your website address. You see what AI tools can read on your site and what to fix first. It is free and runs right here.
Copy this brand file for AI agents.
Save it as plain text or Markdown. Every tool can read that.
brand-rules-for-agents.md
# [Company] brand rules for AI agents Version [date]. Owner: [name]. If these rules and a task disagree, follow these rules. ## Who we are [Company] [does what] for [who]. Use this sentence as written. ## Approved facts (use only these) - [fact] (source: [link]) - [fact] (source: [link]) If a fact is not on this list, do not use it. Ask. ## Voice - [rule 1]. Right: "[line]". Wrong: "[line]". - [rule 2]. Right: "[line]". Wrong: "[line]". - [rule 3]. Right: "[line]". Wrong: "[line]". ## Never - Never use these words: [list] - Never claim: [list, e.g. "first", "best", guarantees] - Never mention prices, discounts or dates that are not approved. - Never name a customer that is not on the approved list. ## Stop and ask a person when - A task needs a fact that is not listed above - The text goes public for the first time on a new channel - Someone complains, asks about money, or asks a legal question - You are not sure ## Visual rules (for agents that make images or pages) - Colours: [HEX codes and where each goes] - Fonts: [names] - Logo: use only files from [link]. Never redraw it.
For the voice part, borrow ready lines from tone of voice prompts. For the facts part, see how my own facts page is laid out.
Brand OS Portal
Get 3 templates in your brand. Free.
Leave your email. I design them by hand and send them to you within 48 hours. It costs you nothing.
Got it. I start on yours today. Your 3 free templates reach you within 48 hours.
I write to you from [email protected]. If you do not see it, check spam.
Make the rules work in any tool.
Tools change often. Your brand should not change with them.
- One file, one source. Every agent loads the same file, not a copy someone edited
- Plain text. No tables in images, no PDF-only rules
- A date at the top. So you know which version an agent used
- Short. Rules an agent can follow in every task, not a book
- Examples beat adjectives. "Right" and "wrong" lines teach more than "friendly but bold"
- Test it. Give two different tools the same task and compare the results
Moving from one tool to another? Read custom GPT and your brand voice and brand voice in ChatGPT and Claude.
Five tests before an agent works alone.
Run these on a copy of a real task. Fix the rules, not just the output.
Motion Posts
Try one Brand OS Portal tool
Rules in, on-brand work out.
This is a real tool from a real Brand OS Portal, on the MTG brand. Type a line. The brand rules shape the video.
Straight answers.
What are brand guidelines for AI agents?
A short plain-text file that tells an agent who you are, which facts it may use, how to sound, what never to say and when to ask a person.
How are agent rules different from a ChatGPT prompt?
A prompt helps one chat. Agent rules also set limits, because an agent may post or send without a person reading it first.
What file format should AI brand guidelines use?
Plain text or Markdown. Every AI tool can read it, and people can read it too.
How do I stop an AI agent from making false claims?
Give it a list of approved facts, tell it to use only those, and tell it to stop and ask when a task needs anything else.
Can one brand file work in Claude, ChatGPT and other tools?
Yes. Keep it in plain words, keep one master copy, and load the same file everywhere.
Where should the brand file live?
In one place the whole team and every tool use, such as a brand portal. Not in five private chats.
Does AI get your brand right? Check it free.
See what ChatGPT says about your company, and how to fix it.
Rather write to me?
Got it.
You hear from me within 2 to 8 hours.