Most AI copy sounds the same because most people never teach the tool who they are. They open a blank chat, paste a request and get back the average of the internet. The fix is a voice file: one document the model reads before it writes a word for you.
In our guide to GPT prompts for brand development we called training your AI an ongoing commitment. This post is the how. It is the same process we use before we write a line of web copy for a biotech or medtech client, and you can build it in an afternoon.
Start with what you already have
You do not need a brand book to do this. You need five to ten pieces of writing that sound like you on a good day. Pull from:
- The best email you sent a customer this year
- Your homepage hero and one service page
- A LinkedIn post that got replies, not just likes
- A proposal or deck intro you were proud of
- Anything a customer quoted back to you
Now pull three pieces that sound wrong. An old press release. A page an agency wrote that never felt like you. The negative examples matter as much as the positive ones. The model learns faster from “never this” than from “more of that”.
Build the voice file
Open a document and write these six sections. Keep each one short. You are briefing a smart new hire, not writing a thesis.
1. Who we talk to. One paragraph. Name the sector and the role. “We talk to founders and marketing directors at Series A to C biotech companies. They are busy, technical and allergic to fluff.”
2. How we sound. Three to five adjectives, each with a sentence of explanation. “Direct: we lead with the outcome and explain after. Warm: we talk to one person, not an audience.”
3. Rules we follow. The mechanical stuff. Sentence case headings. “You” in every paragraph. Numerals for quantities. Short sentences with the occasional long one.
4. Words we never use. This list does more work than anything else in the file. Ours includes navigate, harness, unlock, cutting-edge, seamless, elevate, leverage, robust, landscape and journey. Add the words your industry overuses. In life sciences that usually means innovative, groundbreaking and solutions.
5. Patterns we strip out. Triads (“clear, compelling and credible”). “Not just X, it’s Y.” Openers that describe the world before the reader. Rhetorical questions followed by the answer. Em dashes. Every paragraph the same length.
6. Examples. Paste your five good pieces under a heading that says “Sounds like us” and your three bad ones under “Does not sound like us”. Add one line on each bad example saying why.
Load it the right way
Where you put the file depends on the tool.
In ChatGPT, build a custom GPT and paste the voice file into its instructions. Upload the examples as knowledge files. Name it after your brand so the whole team uses the same one.
In Claude, create a Project and paste the voice file into the project instructions. Upload the examples as project knowledge. Every chat inside that project starts with your voice loaded.
In Gemini, use a Gem the same way.
If you are on a free plan with none of these, paste the voice file at the top of each conversation. It works. It is just more copying.
Test it before you trust it
Run the same three prompts in a fresh chat with the voice file loaded and compare the output to your examples.
Write a 60 word intro paragraph for our homepage. Who we help, what goes wrong without us, what we do.
Rewrite this LinkedIn post in our voice. Keep every fact. Cut it by a third. [paste a post]
Read this draft and list every word or pattern that breaks our rules. Do not rewrite it yet. [paste a draft]
The third prompt is the one people skip and the one that pays off most. Asking the model to audit before it edits shows you whether it understood the rules or just nodded along.
Keep it current
A voice file goes stale in about a quarter. Every time you catch the model producing something that makes you wince, add the pattern to section 5. Every time a piece of copy lands well, add it to the examples. Ten minutes a month keeps the file sharp.
And when the model gets it wrong, say so in the chat the way you would to a writer. “Too formal. We would never say utilize.” The correction improves the next output and tells you what to add to the file.
What this does not replace
A voice file gets you consistent first drafts. It does not give you positioning, and it will not tell you whether your brand should sound that way in the first place. If the voice you are documenting is one you inherited rather than chose, the better move is a brand strategy session before you train anything.
That is the work we do with science and technology companies: define the voice, then build the tools so your team can use it without us. If you want help with either half, start a project. And if you just want more prompts, the original guide has four more.
Want the short version? The 12-prompt brand library has the three prompts from this post and nine more.