The short version: Generic AI prompts produce generic AI writing because they're built from nobody's actual voice, they're built from an average of everyone's. If you want AI output that sounds like you, you have to feed it your own back catalogue, including the messy bits, not a tidy tone-of-voice document. I did this with five years of my own content and it changed how fast I could write, and how my readers reacted.
The problem with "write in a friendly, professional tone"
Every small business owner I talk to has typed some version of that prompt into ChatGPT. Write this in a warm, professional tone. Make it sound conversational but authoritative. And every single time, the output reads like it was written by the same invisible marketing person, because in a sense it was. That prompt is trained on millions of pieces of "warm, professional" content, which means it lands right in the middle. Safe. Smooth. Forgettable.
You can feel it when you read it. It's the reason so much LinkedIn content now sounds identical, three-line hook, bold statement, bullet list, "the truth is..." close. It's not that people are lazy. It's that they gave the AI nothing to work with except a tone label, and a tone label isn't a voice.
What I did instead, and why I nearly didn't bother
When I started rebuilding my business publicly again, after five rough years where speaking gigs dried up and a lot of my old income streams just stopped working, I tried using ChatGPT for first drafts of blog posts and client emails. The drafts were fine. Fine is the problem. My clients don't hire me because my writing is fine, they hire me because I've been doing this for over fifteen years, I've made expensive mistakes, and I say things plainly that other consultants soften.
So I stopped prompting for tone and started building what I now call a voice library. I pulled together roughly 200,000 words of my own past writing: old newsletters going back to 2019, blog posts, LinkedIn updates, even client emails I had permission to reuse. I dumped chunks of it into a document, not as examples of "good marketing writing" but as raw material for the AI to learn my sentence rhythm, my opinions, my specific phrases.
The five-step method for building your own version
You don't need 200,000 words to start. Even 20 or 30 pieces of your own writing will do it. Here's the process:
- Step 1: Pull your archive. Old newsletters, blog posts, social captions, sales emails, even WhatsApp messages to clients if they're substantial. Anything you wrote yourself, unedited by someone else.
- Step 2: Cut it into short chunks. 200 to 400 words each. Long documents confuse the model, short focused chunks teach it faster.
- Step 3: Tag each chunk by type. "Rant," "how-to," "story," "sales email," "objection handling." This matters because your voice in a rant is different from your voice explaining a process.
- Step 4: Write an anti-voice list. A short list of phrases and structures you never use. Mine includes "in today's fast-paced world," any sentence starting with "as an AI," and forced three-item lists where two would do.
- Step 5: Build a living prompt document. Paste 3 to 5 of your best chunks plus the anti-voice list at the start of every new chat, then ask for the actual task. Update it every few months as your writing changes.
Why the polished stuff doesn't work
Here's the part most people get backwards. When they build a "brand voice" document, they pull their best writing, the award-winning case study, the perfectly edited sales page, the piece a copywriter tidied up. That's exactly the wrong material. Polished writing has already had the personality edited out of it, that's what editing does, it smooths.
Your actual voice lives in the throwaway lines. The parenthetical aside you added at 11pm. The blunt one-liner in an email to a client that you'd never put on your website. The typo you left in a LinkedIn post because you were annoyed and hit publish anyway. I fed ChatGPT some of my most unpolished writing, including a few rant-y newsletters I nearly didn't send because they felt too blunt, and that's the material it learned the most from. It's uncomfortable to hand an AI your rough drafts instead of your best work, but the rough drafts are where you sound like a person.
What changed once I did this
Concretely: my average time to draft a 1,500 word blog post went from around 3 hours to under 50 minutes, first draft to something I'd publish with light edits. That's not the AI doing the thinking for me, it's the AI removing the "make it sound like me" step, which used to be the slowest part.
The bigger shift was in reader response. Before, when I used AI-assisted drafts built from generic prompts, I'd occasionally get comments like "well written" or "informative." After building the voice library, I started getting "this sounds exactly like you" and "I could hear your voice reading this." That second reaction is the one that matters for a personal brand, because it means readers trust the words are mine, not a template with my name on it.
Keeping the library from going stale
A voice library isn't a one-off document, it decays. Your opinions shift, the phrases you use change, your business changes. I refresh mine roughly every quarter, pulling in the last few months of writing and quietly retiring older chunks that no longer sound like where I am now. If you skip this, you end up sounding like an older, flatter version of yourself, which is almost as bad as sounding like nobody at all.
This connects to something bigger than blog posts. If you look at how consistent brands feel across every channel, the pattern holds well beyond AI prompting. Look at how Airbnb built a brand around a specific founder story and kept repeating it everywhere, or how a strong Instagram strategy depends on a recognisable voice, not just good visuals. Voice consistency is the whole game, and it's just as true for a solo consultant writing newsletters as it is for a billion-dollar brand.
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.
Where this fits with the rest of your marketing stack
Your voice library isn't only for blog posts. Once you've built it, use the same chunks to prompt AI for email sequences, the kind of thing you'd normally build in a platform like Mailchimp, or for social captions, or for the copy on a landing page. I've even used mine to draft the copy around interactive calculators and tools I've built into pages, because the explanatory text around a tool still needs to sound like a person, not a manual.
Design tools work the same way, incidentally. The way Figma built a brand around a consistent visual language is the visual version of what a voice library does for words: repetition of specific, recognisable choices until they become instantly identifiable as yours. And if you want a reminder of why differentiation beats "professional and polished" every time, go back and read the business lessons from Alex Hormozi, blunt and repetitive beats smooth and forgettable, in marketing as much as in offers.
When it's worth getting help with this
You can build a basic voice library yourself in an afternoon with the five steps above. Where people get stuck is scaling it across a team, or setting it up inside proper workflows so it's not just a document sitting in Google Docs that everyone forgets to open. That's exactly the kind of setup work an AI implementation coach does, building the prompts, the tagging system, and the habit of using it into how your business runs day to day, rather than as a one-time experiment.
Free resource: The Founder-Led Ad Hook Prompt Pack.
Frequently asked questions
How much of my own content do I need before this works?
Twenty to thirty pieces is enough to start noticing a difference, ideally a mix of formats, blog posts, emails, and social captions, because your voice shifts slightly between them. More helps, but quality of selection matters more than volume.
Should I only use my best writing in the voice library?
No, and this is the part people get wrong. Polished writing has had personality edited out of it. Include rants, asides, and unfinished-sounding pieces alongside your tidier work, that's where your actual sentence rhythm and opinions show up most clearly.
How is this different from a brand tone-of-voice guide?
A tone-of-voice guide describes your voice in adjectives, friendly, direct, warm. A voice library is made of actual sentences you wrote, which an AI model can learn patterns from far better than it can learn from a list of adjectives.
How often should I update the voice library?
Roughly every quarter. Add recent writing, remove older pieces that no longer reflect how you write or what you believe now. Skipping this means your AI output slowly starts sounding like an outdated version of you.
Related reading: Why "Write Like Me" AI Prompts Fail for Small Business Content (And What Works Instead) and The AI Receptionist Test Every Small Business Owner Should Run Before Launch.
