How to build a brand voice guide your AI actually follows

Most brand voice guides are written for people, and people fill gaps from context. “Warm, direct, occasionally funny” works on a new hire because they read your old posts and absorb the rest. A language model has no such context, and those three words describe roughly what it already aims for. Paste them into a prompt and the output barely moves. A guide that works for AI is a set of rules specific enough to be broken, plus examples, plus a list of things you never say. This walkthrough builds one in about two hours.

Why doesn’t “friendly but professional” change AI output?

Because the model is already trying to be friendly and professional. Instructions only change output when they constrain it in a way you could check: “keep sentences under 22 words”, “always use contractions”, “never open with a question”. If a line in your guide could be ignored without you being able to tell, it isn’t doing anything.

Step 1: Choose your samples

Pick three to five pieces that sound most like you. Anthropic’s prompting guidance recommends three to five examples, and says they should be relevant to the task and varied enough that the model doesn’t copy the wrong things. Both points apply to voice:

• Same format. Newsletters, scripts and sales pages have different rhythms. Build one sample set per format you write often.
• Your voice, not your best results. The piece that went viral may have been edited by someone else or written in a hurry.
• Varied topics. If all five samples are about pricing, the model learns your topic as well as your style.

If you don’t have three pieces you’re happy with, write one carefully without AI and start from that.

Step 2: Extract the patterns

Analysis is a job AI does well, because it’s counting and describing rather than inventing. Paste or upload the samples and use this prompt:

Analyse these writing samples and describe the author’s style as mechanical rules another writer could follow. Cover: typical and maximum sentence length, paragraph length, contractions, how pieces open, how they close, use of “I”, “we” and “you”, questions, lists, punctuation habits, vocabulary level, jargon, humour, and any words or constructions that recur. Give numbers where you can. Do not use adjectives such as “friendly”, “engaging” or “authentic”. Quote a short example from the samples for each rule.

The quoted examples matter. They let you check each rule against your actual text, and they catch rules the model made up. Spot-check two or three numbers yourself: count the words in a few sentences, look at how your last five pieces actually open.

The analysis will often surface habits you hadn’t noticed, such as a word you overuse or a structure you repeat. Decide whether each one is part of your voice or a tic to remove.

Step 3: Write the guide

Turn the analysis into a document with fixed sections. Here is a template to copy:

VOICE GUIDE: [brand or format]

WHO WE WRITE FOR
[One paragraph: who reads this, what they already know, what they want from us.]

MECHANICS
- Sentences: average 12-18 words, never over 25.
- Paragraphs: 1-3 sentences.
- Contractions: always (it's, don't, we'd).
- Spelling: [British / American].
- Lists: only for 3+ parallel items; otherwise prose.

STRUCTURE
- Open with: [a claim / a concrete situation / the answer]. Never with a question.
- Headings: [questions or statements].
- Close with: [a specific next step / a one-line opinion]. No summary paragraph.

STANCE
- Person: "we" for opinions, "you" for the reader.
- Claims: state them plainly; hedge only when evidence is weak, and say why.
- Jargon: explain any term a beginner wouldn't know, once, in plain words.

TONE BY SITUATION
- Teaching: patient, step by step, examples first.
- Selling: specific benefits, real numbers, no urgency tricks.
- Bad news or apologies: plain, brief, no jokes.

NEVER USE (with replacements)
- "In today's fast-paced world" -> start with the specific situation.
- "Let's dive in" -> go straight to the first point.
- "Game-changer" -> say what changed and by how much.
- Exclamation marks -> a full stop.
- Rhetorical question openers -> a direct statement.

FACTS
Use only facts from the brief or attached files. If something is missing,
write [CHECK] instead of guessing.

The “tone by situation” section borrows a distinction from Mailchimp’s content style guide, one of the best-known public examples: voice stays the same, while tone changes with the situation and the reader’s state of mind. Mailchimp also puts clarity ahead of entertainment, which is a useful rule to copy.

Step 4: Build the do-not list properly

The do-not list usually improves output fastest, because a lot of what reads as generic AI writing comes down to a recognisable set of phrases. Build it by reading AI drafts and noting every phrase that made you wince.

Pair each ban with a replacement. Anthropic’s guidance on formatting is to tell the model what to do rather than only what not to do, and that applies to style too. “No rhetorical questions” leaves a gap; “open with a direct statement of the problem” fills it.

Keep the list to things that actually appear in drafts. Twenty targeted bans beat a hundred speculative ones.

Step 5: Test it blind

A voice guide is a hypothesis until it’s tested.

1. Take a piece you wrote recently that isn’t in your sample set.
2. Write the brief you’d have needed for it: topic, audience, key points, length.
3. Give the AI the brief, the guide and the samples, and ask for the piece.
4. Put your version and the AI version side by side, unlabelled, and read them a day later.
5. Mark every sentence that doesn’t sound like you. For each one, ask which rule would have prevented it, and add or tighten that rule.

To audit drafts quickly, ask a fresh chat to check them:

Here is a voice guide and a draft. List every sentence that breaks a rule in the guide, name the rule, and suggest a rewrite that follows it. Do not comment on anything else.

Run two or three rounds and the guide stops changing much. After that, add rules only when a new failure shows up.

Step 6: Store it where it gets used

The guide only works if it’s applied when you’re rushed, which is when you’re least likely to paste it.

• ChatGPT: create a project for the brand, put the guide in the project instructions and upload the samples as files. Project instructions override your general custom instructions inside the project.
• Claude: do the same with a Claude project, or turn the guide and samples into a skill. Anthropic is moving its older styles feature into skills, and skills need code execution switched on in settings.
• Jasper: upload the samples to a Brand Voice. Jasper Pro includes two Brand Voices; the Style Guide for banned words and terminology is Business-only, so on Pro the do-not list goes into your prompts or the voice description. Our Jasper review covers the plan limits.

Whichever tool you use, keep the master copy as a plain text file you own. It moves between tools, and anyone you hire later gets a usable document instead of a mood board. If you’re still choosing a tool, ChatGPT vs Claude for content compares how each handles saved instructions.

Keeping it current

Review the guide when something changes: a new format, a new audience, a model update that brings new habits. Add to the do-not list whenever a phrase keeps coming back. Retire rules that no draft breaks anymore, because a shorter guide is easier for both you and the model to follow.

My take

Two hours on this does more for AI-assisted writing than switching tools. The step that matters most is replacing every adjective with a rule you could check, then testing it against your own work and fixing failures one rule at a time. Keep the master copy in plain text, store a working copy in the project or skill you write in, and keep growing the do-not list. That list does a lot of the work of making output sound like you rather than like everyone else using the same model.

How we researched this

This guide was built from vendor documentation on prompting and saved instructions, and from a well-known public style guide, checked in November 2026. It is a researched method, not the result of a long-term test, and the template is a starting point to adapt.

• Anthropic prompting best practices: number and type of examples, telling the model what to do instead of what not to do.
• Mailchimp content style guide: the voice versus tone distinction.
• Projects in ChatGPT and Claude skills: where saved instructions live.
• Jasper Brand Voice: how Jasper stores a voice.