GuideNovember 12, 20267 min readTracee team

Brand voice rules an AI will actually follow, with a template

Write brand rules an AI can apply: testable tone, a banned-phrase list a checker can match, 3 to 5 example posts, per-channel notes and a weekly review loop.

Most brand guidelines were written for people, who can infer a lot from "confident but approachable". An AI writing 40 posts a week can't. This guide shows how to turn a brand voice into rules a model can apply and a checker can enforce, with a template you can paste into any AI tool.

Brand voice rules for AI are written instructions that a model reads before drafting and that a check applies after: tone expressed as observable habits, an exact list of banned phrases, a few approved examples, and notes per channel. They work when every rule can be tested against a draft with a yes or no.

Why do AI-written posts drift off brand?

AI drafts drift because most brand guidelines describe a feeling instead of a behavior. "Bold, human, witty" gives a model nothing to check its own output against, so it falls back to its default register, which is the generic marketing voice readers now recognize.

Four causes account for most drift:

Adjectives instead of habits. "Friendly" can mean exclamation marks or it can mean plain words. The model picks one.
Rules phrased only as don'ts. Anthropic's prompting guide advises to "Tell Claude what to do instead of what not to do" (Claude prompting best practices). A list of 30 don'ts with no positive pattern leaves the model guessing what to write instead.
No examples. The same guide calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure."
No check after writing. A rule the model read is a suggestion. A rule a checker enforces is a rule.

How do you describe tone so an AI can apply it?

Describe tone as habits you could spot in a sentence. Mailchimp's public style guide draws the useful line: "Our voice doesn't change much from day to day, but our tone changes all the time" (Mailchimp Content Style Guide). Voice is the constant part, so write it as fixed habits. Tone shifts by channel and situation, so put those shifts in channel notes.

Turn each adjective into something testable:

Vague trait	Rule an AI can apply
Confident	State the claim first. No "we think" or "might". Back it with one number or one example.
Approachable	Second person. Sentences under 20 words. No acronyms without the full term once.
Expert	Name the specific tool, setting or step. Skip definitions the reader already knows.
Witty	At most one light line per post, never at the reader's expense. No puns in headlines.
Humble	Say what didn't work alongside what did. No superlatives about ourselves.

Keep the tone section to five to eight lines. Each extra line competes with the others for the model's attention.

What belongs on a banned phrases list?

A banned phrases list is the set of exact words and phrases that must never appear in a published post. It is the most enforceable part of your brand rules, because checking for a phrase needs no judgment.

Good candidates:

Hype vocabulary: world-class, next-level, disruptive, rockstar, "thrilled to announce".
Filler openers: "Ever wondered", "Picture this", "Buckle up".
Claims you can't support: "industry-first", "guaranteed", "the only platform".
Words with legal or compliance weight in your industry.
Names you don't mention, such as a competitor or a former product name.

Write the list for a literal checker, since most checks (including Tracee's) look for the phrase as written, ignoring case:

Include variants. Add both "world-class" and "world class". The checker won't infer one from the other.
Avoid short entries. A banned "AI" would match inside "said" and "email". Ban the full phrase you dislike, such as "AI-powered".
Ban phrases, not ideas. "Don't overpromise" is a tone rule. "Guaranteed results" is a banned phrase.
Keep it under 40 entries. A longer list usually means some tone rules are hiding in it.

How many examples should you give, and which ones?

Give three to five approved posts per channel. Anthropic's guide recommends 3 to 5 examples "for best results" and says examples should be relevant and diverse (Claude prompting best practices).

Pick examples that:

A reviewer approved without edits.
Cover different formats: a short opinion, a how-to, an announcement, a customer story.
Performed at or above your median for that channel.

Under each example, add one line saying why it works ("opens with the number", "ends with a question to practitioners"). The model copies the pattern you name more reliably than the one it has to infer. If one failure keeps recurring, add a single bad example with its rewrite, clearly labeled as the version to avoid.

What should per-channel notes say?

Per-channel notes describe how tone shifts on each network, in two to four lines each. They shouldn't repeat the voice section.

LinkedIn: First line must stand alone before "see more". Practitioner detail. One idea per post. Link in the post body.
X: One claim per post, 280 characters. No hashtags unless the event has one.
Instagram: Caption supports the image; first sentence names what's in it. Link goes in bio or story, never "link below".
TikTok and YouTube Shorts: Write the hook line that appears on screen in the first two seconds, then the caption.
Reddit: No marketing tone. Answer the thread's question first; disclose that you work at the company.
Google Business Profile: Plain local facts: what's new, hours, offer, one call to action.

Channel notes are also where you set limits the network enforces, so the AI doesn't write a 600-character post for X.

A copyable brand rules template

Paste this into your AI tool's instructions, a project file, or your publishing tool's brand settings, and fill in the angle brackets.

BRAND RULES: <Company>

WHO WE WRITE FOR
<Role, company size, what they already know, what they want from us.>

VOICE (always)
- Second person. Sentences under 20 words.
- Claim first, then one number or one example that supports it.
- Name the specific tool, setting or step.
- No superlatives about ourselves.
- <your habit>
- <your habit>

NEVER SAY (exact phrases, any case)
world-class; world class; next-level; disruptive; thrilled to announce;
ever wondered; picture this; guaranteed; industry-first; <competitor name>

CHANNEL NOTES
LinkedIn: <2-4 lines>
X: <2-4 lines>
Instagram: <2-4 lines>
<other channels>

EXAMPLES (approved, unedited)
1. <post text>
   Why it works: <one line>
2. <post text>
   Why it works: <one line>
3. <post text>
   Why it works: <one line>

BEFORE YOU FINISH
Re-read the draft against NEVER SAY and VOICE. Rewrite any line that breaks a rule.

The final block asks the model to check itself. It helps, but treat it as a first pass, not the enforcement.

How do you run a review loop that improves the rules?

A review loop turns what reviewers keep fixing into rules, so the same edit isn't made 50 times. Run it weekly for the first month.

Collect edits. For every AI draft a person changed before approving, keep the before and after.
Group them. Tag each edit: banned word, tone, length, wrong claim, wrong channel style.
Promote repeats. An edit made three or more times becomes a rule: a new banned phrase, a sharper voice line, or a better example.
Retire dead rules. A banned phrase nobody has triggered in two months can go.
Track two numbers. The share of drafts approved without edits, and the share flagged by the checker. Both should improve month over month.

The approval workflow guide covers how to route drafts so these edits are captured, and the guide on LinkedIn posts that don't read as AI-written goes deeper on one network.

How does Tracee apply brand rules to AI-written posts?

Tracee stores brand rules per workspace in three fields: Tone, a "Never say" list of banned phrases, and Notes for anything else every draft should follow. AI agents connected over MCP, and Tracey (the assistant inside Tracee), read these rules before writing.

Every version of a post is then checked. A version that uses a banned phrase is flagged and can't go out until it's fixed, and the check covers every part of a thread, not only the first post. The match ignores case, which is why the template above lists variants. Each channel's version is also held to that network's rules, such as 280 characters on X or a required video on TikTok and YouTube.

Agents can't change brand rules or guardrails; only people can, and those edits land in the audit log. Brand learning reads your website and past posts to build a brand profile, including words your brand steers clear of and how you post today (themes, habits and what worked), which agents read alongside the rules. The features page shows the brand rules screen, and connecting an AI agent takes one link and a sign-in.

Next step

Take your current brand guide and rewrite its tone section as five testable habits, then list the ten phrases your team deletes most often. Add three approved posts with a "why it works" line. That one page will do more for AI drafts than a 20-page brand book, whichever tool you use; Tracee can hold it and check every draft against it (plans at pricing, at the time of writing, November 2026).

Frequently asked questions

How do I get ChatGPT or Claude to write in my brand voice?

Give it rules it can check against its own draft: concrete habits (sentence length, person, numbers over adjectives), an exact list of phrases to avoid, and three to five real posts you approved, each with a line on why it works. Anthropic's prompting guide calls examples one of the most reliable ways to steer tone and recommends 3 to 5 of them.

What should go in a banned phrases list?

Phrases your team keeps deleting from drafts: hype words, filler openers, claims you can't support, competitor names you don't mention, and words with legal weight such as guaranteed. Write each as the exact text to catch, include common variants, and avoid very short entries that would match inside other words.

How many example posts should brand guidelines include for AI?

Three to five per channel is a practical number. Anthropic's prompting documentation recommends 3 to 5 examples for best results and says they should be relevant and varied. Pick posts that performed and that a reviewer approved without edits, cover different formats, and add one sentence on what makes each one right.

Can an AI enforce banned phrases on its own?

It can try, but a model reading a rule is not a guarantee. The reliable approach is a check that runs on every draft after it's written and blocks publishing when a banned phrase appears. That catches both the model's slips and a teammate's. In Tracee, a version that uses a banned phrase is flagged and can't go out until fixed.

How often should brand voice rules be updated?

Review them weekly for the first month, then monthly. Each time, look at what reviewers changed in AI drafts before approving. Any edit made three or more times becomes a rule: a new banned phrase, a sharper tone line, or a better example. Remove rules nobody triggers, since long lists dilute the ones that matter.