Ali Fakhar
فارسیGet in touch

How to train AI to write in your brand voice

  • Level: Beginner
  • A few hours to set up

Ali Fakhar · Updated: Sept 2026

The short answer

Collect writing your team agrees sounds right, ask a model to turn it into concrete voice rules with examples, edit the rules, and save them as reusable instructions in a Claude Project, custom GPT or Gemini Gem. Test on a real brief and compare drafts blind before anyone relies on it.

On this page

At a glance

The marketing problem
Most AI drafts sound generic because the model is given adjectives such as “friendly but professional”. Rules drawn from real examples, with words to use and avoid, give it something specific to follow and give editors a standard to check drafts against.
Who this is for
Content, brand and social teams, and anyone who briefs AI tools to draft copy.

What you need

Before you start

  • Agreement, from more than one person, on which existing pieces sound right
  • An AI tool that keeps saved instructions (Claude Projects, custom GPTs, Gemini Gems or your company’s approved tool)
  • An owner who keeps the guide current

What goes in

  • Ten to twenty examples across formats: web copy, emails, social posts, a proposal
  • Any existing brand guidelines
  • A few examples that sound wrong, with a note on why

How it works

  1. 01 · You decideCollect your best writing
  2. 02 · AI stepDraft the voice guide
  3. 03 · You decideEdit the guide
  4. 04 · ToolSave it as reusable instructions
  5. 05 · AI stepTest on a real brief
  6. 06 · You decideCompare blind and refine
  • Tool
  • AI step
  • You decide

Step by step

  1. 01 You decide

    Collect your best writing

    Pick pieces the team agrees sound like you, across different formats. Add a few that sound wrong and note why. Remove anything confidential.

  2. 02 AI step

    Draft the voice guide

    Ask the model to describe the voice as rules with a do and a don’t for each, taken from your examples, not as adjectives.

    Prompt to copy

    Below are examples of our writing that sound right, and some that do not. Write a voice guide a new writer could follow. Include: 1. Five to eight rules, each with a do and a don’t taken from the examples 2. Words and phrases we use, and ones we avoid 3. Sentence length, reading level and how we address the reader 4. How the voice changes between [formats, for example web page, email, LinkedIn post] Base every rule on the examples. Do not use adjectives like “friendly” or “professional” without showing what they mean in a sentence. Examples that sound right: [paste] Examples that sound wrong, with notes: [paste]

  3. 03 You decide

    Edit the guide

    Remove rules that describe one writer’s habits rather than the brand, and add anything your guidelines require. Keep it to one or two pages.

  4. 04 Tool

    Save it as reusable instructions

    Paste the guide and three or four of the best examples into the instructions of a Claude Project, custom GPT, Gemini Gem or your approved tool, so every draft starts from them.

  5. 05 AI step

    Test on a real brief

    Use a brief the team is working on now, not an easy one written for the test.

    Prompt to copy

    Following the voice guide in your instructions, write [format] about [topic] for [audience]. Key points: [points]. Keep it under [length]. After the draft, list any rule you found hard to apply.

  6. 06 You decide

    Compare blind and refine

    Mix two AI drafts with two approved human pieces and ask colleagues to pick which is which, and why. Turn whatever gives the AI away into new rules.

Where a person decides

An editor approves the voice rules and decides, from a blind comparison, whether drafts are good enough to use.

What to watch for

  • A voice guide reduces generic phrasing; it does not supply facts, opinions or expertise.
  • Models drift back to stock phrases in long outputs, so keep drafts short and check them.
  • Examples from a single writer can turn a brand voice into one person’s habits.

How to tell it is working

  • In a blind test, reviewers struggle to tell AI drafts from approved examples, or can name what gives them away.
  • Editors log recurring fixes and turn them into new rules.
  • Time from brief to approved draft is compared with before.

An illustrative example

Input

Illustrative: fifteen approved pieces from a fictional accounting-software company, plus three drafts the team rejected as “too salesy”.

Output

A two-page guide with rules such as “Name the task, not the feature: ‘close the month in a day’, not ‘automated reconciliation’”, a list of banned phrases and saved instructions in a shared project.

Want this running in your team?

AI marketing advisory. An experienced marketing partner for the decisions your team needs to make.

View service
All AI in marketing workflows ↗
Analytics preferences

Current preference: off