The short version: Feeding ChatGPT or Claude your old blog posts and asking it to “write like me” almost never works, because most business owners can’t name what their voice is beyond vague words like “friendly” or “professional.” The fix isn’t more examples, it’s a structured 45 minute interview with yourself that pulls out your opinions, your sentence rhythm, and the things you refuse to say, then turns that into a document the AI can follow.
The “just feed it your old posts” advice doesn’t hold up
I see this advice everywhere: paste 20 or 30 of your blog posts into ChatGPT, tell it “write in this style,” and off you go. I tried exactly that in early 2026 with my own back catalogue, forty posts, some of them going back to 2018. The output was fine. It was also nothing like me. It had my vocabulary and none of my rhythm. It used the odd word I use, “punchy,” “specific,” “blunt,” but it wrote in long, safe, rounded sentences where I write short, then long, then short again. It hedged where I don’t hedge.
The reason is simple and a bit uncomfortable: the model isn’t learning your voice from those posts, it’s learning your vocabulary. Voice is structure, not word choice. It’s where you put the full stop. It’s whether you tell people the answer first or make them wait for it. Most brand voice guides, including some I’ve written myself in the past, focus on tone words (warm, bold, approachable) when the thing that makes writing sound like a specific human is sentence architecture and opinion, not adjectives.
The uncomfortable bit nobody selling AI tools tells you
Here’s the part that annoys AI software vendors: the reason your team’s AI output sounds generic isn’t the tool. It’s that most business owners have never sat down and defined their own voice in a way that’s usable. Ask a founder “what’s your brand voice” and you’ll get “friendly but professional” nine times out of ten. That phrase describes about 80% of small business content on the internet, so an AI trained on it produces exactly that: 80% of small business content on the internet.
The businesses whose AI content doesn’t sound like everyone else’s aren’t using a better tool. Duolingo’s team didn’t get a chaotic, slightly unhinged social voice from a prompt, they got it from years of deliberately deciding what the brand would and wouldn’t say, and you can see the same deliberate consistency running through the Duolingo marketing strategy. Revolut’s product marketing has a clipped, numbers-first confidence that shows up whether it’s an app notification or a press release, which is worth studying in the Revolut marketing strategy breakdown. That consistency is a decision made once, documented, and then followed, not something an AI invents for you from a handful of blog posts.
What I did instead
After the failed forty-post experiment, I built what I now call a voice file, a single document, about three pages, that I feed into any AI tool before it writes a first draft for me. It took 45 minutes to create and I’ve only had to update it twice in a year. Here’s what’s in it, and how I built each part.
1. The opinion list (10 minutes)
I wrote down 15 things I believe about my industry that other consultants either won’t say or say more softly. Things like “most AI consultants oversell the tool and undersell the change management” or “a chatbot that isn’t monitored for two weeks is worse than no chatbot.” Voice comes from having opinions and stating them plainly. An AI can’t invent your opinions, so if you don’t hand them over, it fills the gap with balanced, say-nothing sentences.
2. The banned words and banned moves list (5 minutes)
I listed words I never use (“use,” “synergy,” “unlock,” “game-changer”) and structural habits I avoid, like starting three paragraphs in a row with “In today’s world.” This matters more than most people think. Telling an AI what to avoid narrows its output far more effectively than telling it what to include, because “include warmth” is vague and “never start a sentence with ‘In today’s fast-paced world'” is a rule it can follow.
3. Five sentences that sound exactly like me, annotated (15 minutes)
I pulled five real sentences from my own writing and, underneath each one, wrote a line explaining why it works: short sentence, then a longer one that complicates it. Opens with the blunt version of the point, not the polite version. Uses a specific number instead of “many” or “a lot.” This annotated example section did more for output quality than the entire forty-post experiment, because it showed the pattern instead of just the words.
4. The “would I say this out loud” test (10 minutes)
Last section: three sentences an AI generated that I rejected, with a one-line note on why. “Too corporate.” “Sounds like a LinkedIn infographic.” “I’d never end a paragraph on a question like that.” Negative examples train the model almost as fast as positive ones, and this is the part most voice guides skip completely.
What changed once I used it
Before the voice file, a first AI draft of a blog post took me about 25 minutes to edit into something I’d publish under my own name, rewriting openings, cutting hedge words, adding my own examples back in. After the voice file, that edit time dropped to roughly 6 to 8 minutes per post, mostly fact-checking and adding a real number or story the AI couldn’t know. That’s not a small saving. Across a year of weekly posts, that’s around 15 hours of editing time back, which is close to two full working days.
The other change was less measurable but more important: the drafts stopped needing a full rewrite of the opening paragraph, which used to be the single biggest time sink. AI models default to a warm-up sentence before the point. My voice file explicitly bans that, so drafts now open with the answer, the way this post does.
Where this matters beyond blog posts
Voice consistency isn’t a blogging problem, it’s a whole-business problem, and it shows up in places people don’t expect. Hotjar built a distinct, slightly cheeky voice that carries from onboarding emails to error messages to the Hotjar marketing strategy behind their content, and that consistency is what makes the brand feel like one company rather than five departments writing separately. Calm did the same thing in the opposite direction, calm, unhurried, almost sparse copy that matches the product, which you can see clearly in the Calm marketing strategy. If you’re using AI for email replies, proposal drafts, social captions, and even interactive tools like the quiz and calculator content that converts visitors, the same voice file should sit behind all of it, otherwise your quiz result copy sounds like a different company from your welcome email.
There’s also a character element to this that goes back further than AI. Dale Carnegie’s whole approach was built on saying things plainly and specifically instead of hiding behind generalities, and some of the business lessons from Dale Carnegie map almost exactly onto what makes AI writing sound human: specific stories over abstract claims, plain statements over hedged ones. A voice file is really just a modern way of writing down what Carnegie was teaching people to do without one.
Where to draw the line on doing this yourself
Building one voice file for your own writing is a 45 minute job you can do alone this afternoon. Building a voice system across a team, five staff members using AI for client emails, proposals, and social posts, all needing to sound like one business rather than five, is a different job. That’s where most small businesses either give up on AI consistency altogether or bring in outside help to build the templates, the training document, and the review process once rather than five separate times. If you’re at that second stage, it’s worth looking at what an AI implementation coach does day to day, because the useful part isn’t the tool selection, it’s building exactly this kind of voice and process document once and training the whole team to use it the same way.
The three mistakes I see most often
- Treating tone words as the whole job. “Friendly, professional, approachable” describes almost every small business on earth and gives an AI nothing to distinguish you.
- Updating the voice file constantly instead of stress testing it once. I update mine twice a year, not every week, because constant tweaking just confuses the pattern.
- Using the same voice file for every format. A LinkedIn caption and a proposal email need different lengths and different levels of formality even from the same person, so I keep two shorter variants for those.
Frequently asked questions
How many example posts does an AI need to sound like me?
Far fewer than most people think. Five well-annotated example sentences, where you explain why each one sounds like you, teach an AI model more about your rhythm than forty unannotated blog posts. Quantity of raw text matters less than the quality of the pattern you point out inside it.
Can ChatGPT or Claude really learn my exact voice?
Neither model learns your voice permanently between sessions unless you’re using a custom GPT or project with saved instructions. What works is feeding a voice file document at the start of every new chat or session, so the model has your rules and examples in front of it each time rather than relying on memory it doesn’t reliably have.
Is it worth paying someone to build this for a small team?
If it’s just you writing, do it yourself in 45 minutes using the steps above. If you have three or more people using AI for client-facing writing, a paid setup usually pays for itself within a few months through reduced editing time and fewer off-brand drafts going out under the company name.
What’s the single biggest sign an AI draft doesn’t sound like a real business voice?
The opening sentence. AI models default to a warm-up line before the actual point, “In today’s competitive market” or “Many businesses struggle with.” A specific voice states the answer or the opinion first and explains afterward. That one habit, opening with the point instead of the warm-up, fixes more generic-sounding AI writing than any other single change.
Related reading: Grammarly Marketing Strategy: How They Built a Brand That Wins and Why Your AI-Written Proposals Are Losing You the Job (Even Good Ones).
Want the complete version? Read where I break down AI marketing.
