The short version: AI writes like nobody in particular because you’re giving it nothing to work with beyond a topic, so the fix isn’t a cleverer prompt, it’s building a written record of how you talk, feeding that in every time, and still editing the output by hand. Do that and AI becomes a fast first draft machine. Skip it and you’ll spend more time fixing the flat, generic version than you would have spent writing it yourself.
Why your AI drafts read like a corporate press release
I remember the exact post. September 2023, I asked ChatGPT to write a LinkedIn update about a client win, gave it three bullet points about what we’d done, and hit generate. What came back opened with “In today’s fast-paced digital landscape” and closed with “the possibilities are truly endless.” I sound like neither of those sentences. I never have.
The tool wasn’t broken. It did exactly what I asked it to do, which was write generic marketing copy about a topic. I hadn’t told it anything about me. No sentence rhythm, no opinions, no examples from my actual client work, nothing about the fact that I say things like “here’s the bit nobody tells you” or start paragraphs with “look.” AI defaults to the flattest, most averaged version of business writing because that’s the training data it’s pulling from. If you want it to sound like you, you have to hand it you first.
Build a voice document before you write a single prompt
This is the step almost everyone skips, and it’s the one that changes the output. I call mine a voice document, and I built the first version in about ninety minutes on a Sunday afternoon.
- Pull 15 to 20 pieces of your own writing. Emails, LinkedIn posts, a talk transcript, a client proposal, whatever exists. The more varied the format, the better.
- Paste chunks into your AI tool and ask it directly: “Read this and describe my sentence length, my favourite words, how I open paragraphs, and where I use humour or bluntness.”
- Write down the patterns it finds, and add your own. Mine include: I use one-line paragraphs on purpose, I avoid hedging words like “perhaps” and “somewhat,” I mention real numbers instead of saying “a lot,” and I’ll say something is rubbish if it’s rubbish.
- List five phrases you’d never say. Mine include “unlock your potential,” “in this day and age,” and “at the end of the day.” Banning phrases matters as much as choosing them.
- Save all of this as a document, roughly 400 to 600 words, that you paste into every content prompt from now on.
That document is the single biggest lever in this entire process. Without it you’re prompting blind. With it, the AI has a reference point every single time, and you stop getting the same generic draft on repeat.
What goes wrong when people skip this
I worked with a coaching client last year who’d been using AI to write her newsletter for four months and hated every result. She was typing “write a newsletter about time management for busy founders” and getting back something that could have come from any productivity blog on the internet. No voice document, no examples of her own writing pasted in, nothing personal. We built her voice document in one session, and her next newsletter draft needed about ten minutes of editing instead of a full rewrite. Same tool, same topic type, completely different result because the input changed.
The prompt structure that works
Once you have the voice document, here’s the prompt structure I use for most content:
- Paste the voice document first, every time.
- State the topic and the one thing you want the reader to walk away knowing.
- Give it one real example or story from your own experience to include, even if it’s rough. AI can polish a rough anecdote far better than it can invent a convincing one.
- Set a word count and tell it to write in short paragraphs, no more than three sentences each.
- Tell it explicitly what to avoid: no rhetorical questions to open paragraphs, no “in conclusion,” no exclamation marks stacked at the end of sentences.
That last instruction matters more than people think. AI models default to a handful of tics: rhetorical questions, “let’s dive in,” lists that repeat the same sentence structure three times in a row. Naming the tics you hate and banning them by name works better than asking for “a natural, conversational tone,” which means nothing to the model.
Where AI still needs you to do the boring work
Here’s the bit that most people selling AI writing courses conveniently leave out: even with a perfect voice document and a well-built prompt, you still have to edit the draft, and that editing pass can take almost as long as writing a short piece from scratch would have taken. AI gets you 70 percent of the way to sounding like you. The last 30 percent, the specific joke, the number you happen to remember from a client call, the sentence that only makes sense because you were in the room, that part is still yours to add by hand every single time.
This is uncomfortable for anyone hoping AI content is a way to stop writing entirely. It isn’t. It’s a way to stop starting from a blank page. Those are very different things, and confusing them is why so much AI-assisted content still reads as hollow even when the person publishing it swears they “used their own voice prompt.” The prompt gets you the skeleton. You still have to put the blood in it.
If you’re trying to work out where that line sits for your own business, our breakdown of how much AI-written content is acceptable for your website goes into the specific ratios I use with clients, because “some AI is fine” isn’t a useful answer on its own.
The edits that make the biggest difference
After hundreds of AI drafts run through this process, these are the five edits I make almost every time, in order:
- Cut the first paragraph entirely. AI almost always writes a throat-clearing intro you don’t need. Start with sentence two or three.
- Break up any paragraph over four sentences. AI writes in even, tidy blocks. Real speech doesn’t.
- Add a number. AI says “significant increase,” I say “up 34 percent in six weeks.” Specificity is the fastest way to sound human.
- Remove every instance of “moreover,” “furthermore,” and “in addition.” Nobody talks like that.
- Read it out loud. If you’d never say a sentence to a colleague over coffee, rewrite it until you would.
That last one takes 90 seconds and catches more fake-sounding content than any other step on this list.
When it’s fine to leave mostly untouched, and when it isn’t
Not every piece of content needs the full voice treatment. A meeting summary, an internal FAQ, a straightforward how-to for your own team, AI can write those close to final draft with light editing, because nobody expects personality from them. This connects to a bigger point about what content marketing means and why it matters for your business: content marketing is meant to build trust through recognisable voice over time, and that only works on the pieces where a reader expects to hear from you specifically. LinkedIn posts, newsletters, opinion pieces, anything with your name on the byline, those need the full process every time. Product descriptions and internal documentation generally don’t.
Email is its own case worth mentioning separately. Newsletter subscribers notice tone shifts faster than almost any other audience because they’re reading you regularly and comparing each send to the last one, which is part of why email marketing is still important for small business in a way that generic blog content sometimes isn’t. If your newsletter suddenly sounds like a stranger wrote it, people notice within two sends, and open rates drop before you’ve even worked out why.
What this means for writers and for hiring
This shift has changed what counts as a content writing job today. The writers doing well right now aren’t the ones typing every word from scratch, they’re the ones who can build a client’s voice document, run the prompts, and do the editing pass that makes it sound like the client rather than like a tool. That’s a different, arguably more valuable skill than pure drafting, and it’s part of why the data on what freelance writers realistically earn shows the ones charging premium rates now are selling judgment and voice work, not word count.
If you’re a business owner trying to sort this out for your own content and you don’t have the time to build the system yourself, this is exactly the kind of thing an outside pair of hands helps with fast. I’ve built voice documents and prompt systems with clients through my AI implementation coaching work, and it’s usually a two or three session job, not a months-long project, because most of the work is extracting what’s already in your head onto a page the AI can read.
A quick worked example
Here’s the difference in practice. Generic prompt: “Write a LinkedIn post about why small businesses should use AI.” Generic output: “In today’s competitive landscape, small businesses that fail to adopt AI risk falling behind. AI offers numerous benefits including increased efficiency and cost savings.” Nobody’s voice, nobody’s numbers, could have been written by anyone about anything.
Same topic with a voice document and a specific instruction to include one real number: “I had a client last month who was spending six hours a week writing social captions by hand. We built her a voice document and a prompt template in one session. Six hours became forty minutes. That’s not AI replacing her, that’s AI doing the bit she never wanted to do anyway.” Same topic, same tool, completely different result because the input included a real story and a real number instead of a topic and a hope.
Frequently asked questions
How long does it take to build a voice document that works?
About 90 minutes for a first working version if you already have 15 to 20 pieces of your own writing to pull from. Plan to update it every few months as your writing changes, it’s not a one-off document you set and forget.
Can AI ever fully replace me writing my own content?
Not for anything with your name on it that’s meant to build trust over time, no. It can replace the blank page and the first draft, but the specific details, the real numbers, and the final edit still need to come from you or someone who knows your voice as well as you do.
What’s the single biggest mistake people make when trying to sound like themselves in AI content?
Skipping the voice document and going straight to prompting. Asking for “a natural, conversational tone” means nothing to the model. Giving it 400 words describing your actual sentence patterns, phrases you’d never say, and real examples of your writing changes the output far more than any clever phrasing in the prompt itself.
How much editing should I expect to do on an AI draft?
With a good voice document in place, expect to cut the first paragraph, break up long blocks of text, add at least one real number, and read the whole thing aloud before publishing. That’s usually a 10 to 15 minute pass on a 600 word piece, not a full rewrite, but it is not optional.
