The short version: Training AI on your brand voice is not about writing a one-line style guide and hoping for the best. It requires feeding the model real examples of your best writing, building a structured voice document with rules and anti-patterns, and testing it repeatedly until the output passes a "could I have written this?" test. Done right, AI becomes a genuine extension of your voice rather than a beige replacement for it.
Why most people get terrible AI output (and blame the AI)
I hear this constantly from clients. "The AI doesn't sound like me." "It's too corporate." "It uses words I would never use." And yes, sometimes that is a model limitation. But nine times out of ten it is a prompting and training problem, not a technology problem.
When you open a blank chat window and type "write a LinkedIn post about my new service," you are giving the AI absolutely nothing to work with. It defaults to the average of every professional LinkedIn post it has ever seen. That average is beige. It is safe. It uses words like "excited to announce" and "thrilled to share" and ends with three questions nobody will answer.
The AI is not broken. You just have not told it who you are.
This post is my real working method for fixing that. I use it for my own content and I have walked dozens of clients through it. It takes a few hours upfront and it saves you enormous amounts of frustration every single week after that.
Step one: build your raw voice library
Before you can train anything, you need source material. This is the step most people skip entirely, and it is why their results stay mediocre.
Go and find 15 to 20 pieces of content you have written that you are proud of. Not content that performed well, necessarily. Content that sounds most like you. That distinction matters. A post might have gone viral because of timing, not voice. You want the pieces where, if someone read them blind, they would say "that is definitely you."
This could be:
- Blog posts or newsletter issues
- LinkedIn posts or Twitter threads
- Email sequences you have written
- Sales page copy
- Podcast transcripts (cleaned up)
- Even long voice notes you have transcribed
Copy all of it into a single document. Do not edit it yet. You want the raw material, including the quirks, the sentence fragments you use deliberately, the specific words you reach for, the rhythm of how you build to a point.
If you are early in your content journey and do not have 15 to 20 pieces, write five short paragraphs right now, stream of consciousness, about what you do and why you do it. That is still workable material. It is just thinner, so your outputs will need more manual correction at first.
Step two: write a proper brand voice document
This is where the real work happens. A brand voice document is not "we are professional but approachable." That tells the AI nothing it does not already know. You need specifics.
Your voice document should include the following sections, and I am going to show you what mine looks like in places.
Tone descriptors with examples
List four to six words that describe your tone. Then, critically, show the AI what each one looks like in practice.
For me, one of my tone words is "blunt." Here is how I show that to the AI:
Blunt, meaning: I say the uncomfortable thing directly. Example of blunt: "Most of your social media content is a waste of time, and you know it." Example of NOT blunt: "It can sometimes be challenging to see results from social media content."
See the difference? You are giving the model a before and after. That is far more useful than a single adjective.
Vocabulary list: words I use and words I never use
This is one of the highest-value things you can do. Spend 20 minutes going through your source material and pull out words and phrases you reach for repeatedly. These are your "voice fingerprint" words.
Then list the words and phrases you never want to see in your content. For me that list includes: "use," "unlock," "game-changer," "thrilled," "exciting times," "in today's fast-paced world," and about 30 others. I update the list every time I catch the AI using something that makes me wince.
Sentence structure patterns
Do you write long, flowing sentences with subordinate clauses? Do you write short punchy ones? Both? In what ratio? Do you use fragments deliberately?
I write in short bursts with occasional longer sentences for context. I use one-line paragraphs as punches. I told the AI exactly that, with examples pulled from my own work.
What you talk about vs. what you do not
List your core topics. Then list the adjacent topics you actively avoid, even if they are popular in your space. This helps the AI stay in your lane rather than drifting into generic industry content.
Audience assumptions
Who are you talking to? What do they already know? What level of jargon is appropriate? I write for business owners and marketers who are not afraid of specifics, so I can reference real tools, real numbers, real case studies without explaining every term from scratch.
Step three: choose where to embed this training
You have two main options and I use both.
Option A: Custom instructions in ChatGPT
If you use ChatGPT, you can paste a condensed version of your voice document into the custom instructions section (Settings, then Personalisation, then Custom Instructions). This applies to every conversation by default. It is not a perfect system because it has a character limit, so you will need to distil your full voice document down to the most critical rules. About 800 to 1200 words fits well.
Option B: A custom GPT
This is the approach I now use for most client work because it gives you far more control. You can upload the full voice document as a file, write detailed system instructions, and share the GPT with your team. I wrote a full step-by-step guide to building a custom GPT for your small business if you want the full technical walkthrough, but the short version is: create a new GPT, paste your full voice document into the instructions, upload your sample content as knowledge files, and write a clear system prompt that tells the model its job.
Option C: A master prompt you copy every time
If you use Claude, Gemini, or another model, or you do not want to build a custom GPT, you can keep a master prompt document that you paste at the start of every session. It is more manual but it works. I used this method for about eight months before I moved to custom GPTs and it produced solid results.
The honest point most articles will not make
Here it is. The uncomfortable truth about AI brand voice training that I have not seen written plainly anywhere else.
Your AI output will only ever be as distinctive as your actual voice.
If your writing is generic to begin with, if you have never developed a real point of view, if you write like every other consultant in your industry, then training the AI will just give you faster generic content. The AI cannot manufacture a voice you have not built.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
This is why I think the exercise of building a voice document is valuable regardless of AI. It forces you to ask: what makes my writing mine? Where do I have opinions? What words do I reach for? A lot of people sit down to write their vocabulary list and realise they cannot fill it. That is important information.
If your voice document feels thin, the answer is not to skip the AI training. The answer is to go and write more. Publish more. Talk out loud and transcribe it. Your voice develops through repetition and publishing, not through planning. Once you have more source material, go back and update the document. Making your brand memorable in this environment requires a real perspective, not just a consistent colour palette.
A real example from my own testing
About 18 months ago I was working with a life coach in the US, let's call her Amanda, who was getting AI content that read like a corporate wellness brochure. Warm but empty. No edge, no specifics, no her.
We spent three hours together building her voice document. The biggest discovery was that Amanda's best writing voice was conversational and slightly irreverent. She swore mildly in her emails. She made fun of therapy-speak. She said things like "no, your limiting beliefs are not the problem, your schedule is." That voice was completely absent from her AI output because she had never told the AI any of it.
We pulled 12 of her best email newsletter issues, identified her vocabulary fingerprint, listed the therapy-speak phrases she specifically hated and wanted to avoid, and built her a custom GPT with about 2,000 words of instructions and those 12 newsletters uploaded as knowledge files.
The output quality difference was significant from the first test. Not perfect. We spent another hour doing what I call "voice correction loops," where she rated each output paragraph and I used those ratings to refine the instructions further. But the gap between "this sounds nothing like me" and "I could send this" closed in one afternoon.
She told me three months later she had gone from spending four hours on a newsletter to spending 45 minutes. That is a real number. Not a promise, a specific result from a specific person with a specific process.
Step four: the testing and correction loop
Training is not a one-time event. Here is the ongoing process I use.
Week one through two: Generate content using your trained system. For every piece, read it aloud and mark every sentence that does not sound like you. Be specific about why. "Too formal" is not useful. "Uses passive voice when I always use active" is useful. "Says 'it is important to' when I would say 'here is why this matters'" is useful.
Week three through four: Take your corrections back into your voice document. Add the specific anti-patterns you caught. Update your vocabulary blacklist. Add new examples to your tone sections.
Month two onwards: Your correction load will drop significantly. You will still need to edit, but the heavy lifting of voice conversion will diminish. At this point you are mostly catching edge cases.
I review my own voice document roughly every quarter. I add new phrases I have started using, remove ones I have moved away from, and update my sample content with newer writing that better represents where I am now. Your voice evolves. Your training should evolve with it.
How this integrates with your wider content system
Brand voice training does not exist in isolation. It connects directly to how you produce content at scale. Once your AI is trained on your voice, the whole promise of AI-assisted content becomes real rather than theoretical.
I have written about how to automate your content calendar with AI as a full system, and voice training is the non-negotiable foundation of that system. Without it, you are automating mediocre content at speed, which is worse than producing less content that sounds human.
The goal is not to remove yourself from the content. The goal is to remove the friction of getting from thought to published piece. When the AI is trained, you are editing and refining rather than rewriting from scratch. That distinction is where the time saving lives.
For businesses where brand voice is the entire competitive advantage, like personal brands, coaches, consultants, service businesses where people are buying you, not just what you do, this is not optional. If your AI content sounds generic, it is actively undermining the thing that makes you worth hiring. Automating your marketing without losing your brand voice is possible, but only if you do this foundational work first.
Quick reference: what to include in your voice document
- Four to six tone descriptors, each with a "sounds like this / does not sound like this" example
- A vocabulary list: 10 to 20 words you use regularly
- A blacklist: words and phrases you never want to see
- Sentence structure description with examples from your own writing
- Paragraph length patterns (how long are your paragraphs typically?)
- Three to five content topics you always address
- Three to five topics you actively avoid
- Audience description: who they are, what they know, what they do not
- Any formatting preferences (do you use subheads? Bullet lists? How often?)
- Examples of your best work, pasted directly in or uploaded as files
Total document length: aim for 1,500 to 3,000 words. Long enough to be specific, short enough that the AI can hold it in context without losing the important bits.
One thing worth noting for service businesses specifically
If you work with clients on their content, whether you are a copywriter, a social media manager, a VA, or a consultant, you can use this exact process to build voice profiles for each client. One custom GPT per client, trained on their best content. It scales your output without scaling your hours, and it means you are producing copy that sounds like them rather than like you filtered through a generic AI tone.
This is now a service I offer as a standalone deliverable. A voice document plus a trained custom GPT, handed over to the client with a walkthrough. It takes about half a day to build and it is one of the most useful things you can give a client who is trying to use AI in their business. You can also apply the same thinking to audio content: if you are using podcasts to build brand awareness, transcribing your best episodes and feeding them into your voice document gives the AI real spoken-voice material to learn from, which is often more authentic than polished written copy.
Frequently asked questions
How long does it take to train AI on your brand voice?
The initial setup, building your voice document, selecting sample content, and creating a custom GPT or master prompt, takes between three and six hours depending on how much source material you have and how detailed you go. The ongoing refinement happens in parallel with your normal content production and adds about 10 to 15 minutes per piece in the first month, dropping significantly after that.
Do I need to rebuild my voice training every time I use a new AI tool?
Not from scratch. Your voice document is model-agnostic. You write it once and paste it into any system. The formatting of how you deliver it (custom GPT vs. system prompt vs. custom instructions) changes by platform, but the content of the document itself transfers directly. Keep it in a Google Doc or Notion page and pull from it whenever you set up a new tool.
What if I do not have a consistent brand voice yet?
Then building the voice document will reveal that gap, and that is useful. Write five to ten pieces of content before you attempt the training process, focusing on being honest and specific rather than polished. Use those raw pieces as your source material. Your voice will not be fully formed yet, but you will have something for the AI to work from, and you can update the training as your voice develops over the following months.
Can AI fully replace a human copywriter who knows your brand?
No, and I say that as someone who uses AI heavily in my own content work. A trained AI will get you to a strong first draft significantly faster than starting from scratch, but it does not have strategic judgment, it does not know what happened in your business last week, and it cannot catch its own off-brand moments. You still need a human in the loop for editing, strategy, and quality control. What changes is the ratio of time spent creating versus refining.
Related reading: How to Set Your Rates as a New Virtual Assistant and The Virtual Assistant Niches That Pay the Most in 2026.
