How to Build an AI Voice Bank From Your Old Content So You Stop Sounding Like Everyone Else's ChatGPT

The short version: If your AI-written content reads like everyone else's AI-written content, the fix isn't a better prompt, it's feeding the model years of your own old writing first so it has something real to copy from. I built one from 63 blog posts and about four years of newsletters, and it took a weekend, not a course.

The problem nobody wants to say out loud

Every LinkedIn post about "AI writing" tells you to write a good prompt. Give it your tone of voice. Tell it three adjectives that describe your brand. I did all of that for about eighteen months and my content still came out sounding like a slightly posh corporate wellness brand, not me.

Here's the bit people skip: a prompt is instructions, not evidence. When you type "write in a warm, direct, no-nonsense tone," you're describing a voice, not providing one. The model still has to guess what "warm and direct" means to you specifically, and it guesses using whatever generic version of "warm and direct" it has seen a million times in training data. That's why so much AI content has the same rhythm now, the same three-word punchy sentences, the same "here's the thing" openers. Everyone is prompting for the same adjectives.

What a voice bank is

A voice bank is a document, or a folder of documents, made entirely of your own real writing, that you feed to an AI tool before you ask it to write anything new. Not instructions about your voice. Your actual voice, in bulk.

Mine is a single 40,000-word file. It contains:

• 63 blog posts from 2020 to 2026, picked because I remembered writing them myself, no AI involved at all
• About 90 newsletter emails, chosen because they got the highest reply rates, not the highest open rates
• 14 transcripts of talks I've given, cleaned up but not rewritten
• A short list of phrases I say and phrases I never say (I do say "right, let's be blunt", I never say "in today's fast-paced digital landscape")

That file goes into a project space (I use Claude's Projects and a custom GPT in parallel, they behave slightly differently and it's useful to compare) as permanent context. Every time I ask it to draft something, it's drafting with 40,000 words of me sitting behind the prompt, not just three adjectives.

The exact steps, no fluff

This took me one Saturday afternoon, roughly four hours, and I've refined it twice since. Here's what I'd tell a client to do this week.

• Step 1: Pull 40 to 80 pieces of your own unedited writing. Old emails to clients, blog posts from before you touched AI, DMs where you explained something to a customer. Volume matters more than polish. I'd rather have 60 scrappy emails than 10 perfect blog posts.
• Step 2: Cut anything written by a ghostwriter or an agency. This sounds obvious but almost every business I've worked with has years of "their" content that a freelancer wrote. If you feed that in, you're building a voice bank of someone else's voice.
• Step 3: Paste it all into one document, unformatted, no headings needed. Aim for 30,000 to 50,000 words if you can. Less than 15,000 and the model doesn't have enough pattern to lock onto.
• Step 4: Load it as persistent context, not a one-off paste into the chat. Every major tool has a version of this now, memory, projects, custom instructions with file attachments. The point is it needs to be there every single time you write, not just the first time.
• Step 5: Write a one-page "anti-style guide" alongside it. List sentence structures, transitions, and words you use, and separately, words you'd never let through. This is where I ban words like "use" and "" for myself, in writing, so the model stops offering them.
• Step 6: Test it blind. Take three new drafts, one written by you, one by AI with no voice bank, one by AI with the voice bank. Send them to five people who know your writing well and ask them to guess which is which without telling them what the test is.

I ran that last step with 40 newsletter subscribers in early 2026. I sent three short "about me" style updates, no context given. 71% correctly picked the voice-bank version as "sounds most like Lilach." Only 34% picked the no-voice-bank AI version correctly, most people guessed it was written by an assistant. That gap is the whole point of doing this.

The uncomfortable truth about voice banks

Here's the part most of the "AI tone of voice" advice out there avoids saying clearly: a voice bank doesn't teach AI to write like you. It teaches AI to average you, and an average of you is still not you.

I noticed this most on my own content. The voice-banked drafts were good, better than generic AI output by a mile, but they were a smoothed version of me. All my consistent tics were there, the short punchy lines, the Britishness, the blunt opinions. What was missing was the mess, the bit where I go off on a tangent about a client call from 2019 and it turns out to be the best line in the whole piece. AI averages toward your patterns, and your best writing usually breaks your own patterns on purpose.

So the voice bank gets you 80% of the way, fast, and useful for drafts, emails, social captions, first passes at blog posts. The last 20%, the sentence that makes someone screenshot the post, still has to come from you sitting down and rewriting one paragraph. Anyone telling you AI plus a voice bank replaces the writer entirely is selling you something.

Where this pays off in a small business

The obvious use is blog posts and newsletters, but the bigger win for most small businesses I work with is customer-facing writing that used to get outsourced or skipped entirely: proposal emails, FAQ pages, onboarding sequences, the "sorry for the delay" email that usually gets written in two rushed minutes. Those pieces of writing carry your brand voice constantly, far more often than your blog does, and they're usually the ones with zero voice consistency because whoever's free writes them.

One client, a five-person accountancy firm in Leeds, built a voice bank from 18 months of their founder's client emails. Within six weeks, every team member was drafting client updates through the same custom GPT, and the founder stopped having to rewrite junior staff emails before they went out, which had been eating about four hours of her week. That's the real ROI, not "better blog posts," but hours back and a consistent voice across people who aren't you.

If your team is small and stretched, this is also exactly the kind of project worth getting outside eyes on early, because the failure mode is building a voice bank from the wrong material and not realising it for months. If you want a second pair of hands setting this up, this is the sort of thing I cover with clients through hands-on AI implementation coaching, rather than a generic course.

How this connects to the rest of your content strategy

A voice bank isn't a stand-alone trick, it works best sitting inside a wider content system. If you're already doing content curation to fill gaps between original posts, feed the curated commentary through the same voice bank too, otherwise your curated content and your original content will sound like two different businesses.

It's also worth checking whether any of this is even reaching AI search results in the first place. There's no point having a beautifully consistent voice if ChatGPT and Perplexity are quietly skipping your site altogether when people ask questions in your niche. Run the 15-minute AEO audit before you spend more hours refining tone, because consistent voice on pages nobody's citing is a lower priority fix.

And if you want to see what "consistent voice at scale" looks like when a much bigger brand does it, look at how Red Bull's marketing strategy stays recognisable across hundreds of pieces of content a year, or how Wix built a distinct brand voice despite being a huge, multi-person marketing operation. Neither of those brands got consistent by writing a tone-of-voice PDF once and hoping everyone read it. They built systems.

A quick honesty check before you start

Before you spend a weekend on this, be honest about one thing: do you have 40,000 words of your own real, unedited writing sitting around? If your last five years of content was all written by an agency, a ghostwriter, or an AI tool from day one, you don't have a voice bank to build yet. You have a voice to find first, and that's a different, slower job, closer to the kind of reflective work in using ChatGPT prompts for self-discovery and reflection than a content project. Give that a fortnight of proper journaling or voice notes transcribed before you try to bank a voice that doesn't exist on paper yet.

Related: the chatgpt page.

Frequently asked questions

How much old content do I need to build an AI voice bank?

Aim for 30,000 to 50,000 words minimum, roughly 40 to 80 pieces of writing. Under 15,000 words, most AI tools don't have enough pattern to lock onto and will default back to generic phrasing.

Which AI tool is best for building a voice bank?

Any tool with persistent project memory or custom instructions with file uploads works, including Claude's Projects and custom GPTs in ChatGPT. The tool matters less than the quality and volume of the writing you feed it.

Will an AI voice bank make my writing sound exactly like me?

No, and be wary of anyone claiming it will. It gets you roughly 80% of the way, a solid, consistent draft that sounds recognisably like you, but the sentences that make a piece memorable still need a human pass.

Can I use this for a whole team, not just myself?

Yes, that's often where it earns its keep most. Build the voice bank from the founder's or lead voice's material, then give the whole team access to the same project so client emails, proposals, and social posts sound consistent no matter who wrote the first draft.

Where to check the details

• OpenAI's own platform documentation
• Anthropic's own documentation