I spent two years trying to get AI to write like me. The breakthrough wasn't a better prompt. It wasn't a fancier model. It was an architecture.
Here's the problem nobody selling AI tools wants to talk about. AI almost writes like you. When the audience is yours and the post has to ship under your name, almost might as well be failing.
The standard playbook is to throw more prompt at the problem. "Write in the voice of [name], who is known for X, Y, Z." You write a prompt, generate a draft, edit it, write another prompt, generate again. Three rounds in, you're rewriting the entire thing by hand because the AI keeps drifting back to its default register — the slick, hedged, AI-trained voice that sounds like every other piece of content on the internet right now.
I went through this loop for months. I tried different models, different prompts, different temperatures, different framings. The output was always close enough to be tempting, always wrong enough to need a full rewrite.
The shift happened when I stopped trying to solve it with prompts and started building infrastructure instead.
What was missing
The thing AI tools won't tell you is that your voice isn't a style you can describe in a prompt. It's a thousand small decisions you make every time you write — what you don't say, what you cut, which words you use in which contexts, which clichés you'd rather die than utter, how you build an argument, how you close.
A prompt can capture maybe ten of those decisions. The other nine hundred and ninety live in your existing writing. They're patterns the AI has to read enough of you to absorb. A description won't teach them; only material can.
The fix is to give it that material — systematically — before it ever drafts.
The architecture
Three layers, built over the last eighteen months.
The first layer is identity. A set of always-loaded rules about who I am, how I work, what TAP Creative is and isn't, which language I've banned and which I keep. This sits at the global level. Any AI tool I'm working with reads these before it does anything else. It's the constitution.
The second layer is feedback. Every time I correct something — "no, don't write it like that" — that correction gets captured as a rule. Not because the model learns from it directly, but because the rule loads into context next time and the model doesn't make the same mistake twice. Over the last year this layer has grown to about forty rules, each one a real correction I made to a real piece of writing, each one explaining not just what to do but why. The why is what lets the model apply the rule to situations the rule didn't explicitly name.
The third layer is the corpus. A hundred-plus files of my actual quotes — voice notes, written corrections, things I've said in sessions about copy and tone and register. Tagged by topic. When I'm about to draft something for a client website, the system pulls the files about positioning, about clarity, about how I think about clients. When I'm about to draft something for a LinkedIn post, it pulls a different set — thought leadership, self, the way I write about my own work. Before any draft starts, the relevant corpus is loaded into context.
That third layer is the one that does the heavy lifting. It's the difference between asking an AI to write in Lola's voice and giving it Lola's voice to read first.
Why this works
Once all three layers are loaded, the model isn't trying to imagine what I sound like. It's been given a thousand pieces of evidence. The drafting that follows is closer to a first pass that needs editing than a first pass that needs rewriting.
It isn't perfect. I still edit. I still cut sentences that aren't quite right and rephrase things that feel borrowed. But the gap between what AI produces and what I would actually ship has shrunk from rewrite the whole thing to edit the seams. That's the difference between AI as a glorified autocomplete and AI as an actual collaborator.
The other thing it does is preserve voice over time. The corpus grows with me. Every correction goes back into the rules. Every session adds quotes. The system doesn't decay — it compounds. A year from now my AI collaborator will know me better than it does today, and the year after that, better still. That isn't true of any prompt-based approach.
What this means for clients
This is the same architecture I'm building now for TAP clients who want thought leadership at scale.
The pitch for most AI-content services is volume — get more posts out, faster, for less money. What I'm building runs the other direction. It's slower up front because you have to build the voice infrastructure properly. But the posts that come out the other end sound like the person they're attributed to. They survive the audience's scepticism. They build the recognition the person actually wants.
If you're a senior professional with a real point of view but no time to write everything yourself, the question isn't how do I get AI to generate more content. It's how do I build the architecture that lets AI draft credibly in my voice.
Building that architecture is the real job. The drafting is the easy part — once the foundations are in.
The honest part
I'm still building this. The system has a name in my head — Lola's Brain — and an architecture I'm refining. The current version is internal infrastructure. The next version is the one I'll package up. Some of you will be the first to try it.
If you're tired of AI content that sounds like everyone else's AI content, this is the architecture worth building. If you want help building it for your own voice, that's the conversation worth having.