Meet Maya
We run into this with customers all the time — let's call this one Maya.
Maya runs a small home-goods shop. She posts three times a week, which already puts her ahead of most, and she uses AI to write the captions because it saves her the better part of an hour each time. The posts are clean, on-topic and on schedule. They also stopped sounding like her a while ago, and her numbers have been flat ever since.
She tried the obvious fixes. She added "warm, conversational tone" to every prompt. She switched tools twice. Every caption still came back polished and interchangeable — the kind of copy any home-goods shop in any city could post.
So one evening she tried something simple. She opened her last ten captions, and next to them ten posts from the two shops she competes with most. Then she covered the names.
She couldn't reliably tell which ones were hers.
That's not a bug, and it isn't her fault. Every time she opened a new prompt, the tool started from nothing — no memory of her product notes, her About page, the captions that had landed with her people. So it filled the gaps with what an average shop in her position would sound like. Average in, average out.
What changed things wasn't a better prompt. She stopped describing her voice and started showing it: her website, her best captions, the blurb she'd written by hand for a market stall. Once the tool was working from her own sentences, the drafts opened the way she actually talks — and her regulars noticed before she said a word.
The part a better prompt can't fix
When AI copy sounds generic, the instinct is a longer prompt: more adjectives, the tagline, "sound friendly but expert." It helps for one post. Then the next session starts from zero again, because nothing carried over.
A blank prompt doesn't produce a bad answer. It produces the safest one — the answer an average, credible business would give, because nothing in the request said who's asking. That's the real reason AI content sounds generic, and a newer, more expensive model gives you the same thing, just more fluently. A tone dropdown doesn't change it either: "confident" or "playful" from a list is still starting from zero, just a differently labelled zero.
Examples beat descriptions. Most people are bad at describing their own voice and very good at having already used it. Your old posts, your site and your product copy already contain the voice. The model only needs to see them before it writes, every time.
Try this week
Run the name-swap test. Take your last three posts. Replace your business name with your closest competitor's. If a post still reads as true and natural, nothing in it was about you.
Then look for one sentence only you could have written — a real detail from your week, a phrase you use with customers, a number from your own work. No such sentence? That post came from the statistical middle.
Fix the input, not the draft. Before your next AI request, paste in three of your own posts that sounded like you, plus two lines on who you write for. Then ask for the draft. It takes two minutes, and you'll see the difference in the first line.
Or skip the routine: this is the step Zolt does for you. It learns your voice once from your website and your own posts, then writes every draft from it — so there's nothing to paste.
From the Zolt blog
Stop re-explaining your business to every AI tool — why pasting context every session wears thin, and what to keep instead.
AI writing sounds like a chatbot? Here's why — if you spend longer fixing drafts than writing them, this is the missing piece.
Nobody reads past the first line of your posts — once the voice is yours, make the opening line earn the rest.
That's it for this week. If you'd rather your posts started from your voice without the copy-and-paste, Get Started.
— Your friends at Zolt
P.S. Try the name-swap test on a competitor's posts too. It's oddly satisfying.