Give ChatGPT a description of how you write and three or four real examples of it, in a place it reads every time: custom instructions for one person, a project or a custom GPT for a team. Then check each draft against that description yourself, because ChatGPT isn’t checking. OpenAI says as much: ChatGPT won’t always interpret custom instructions the way you meant.
The three places ChatGPT can hold your voice
Custom instructions are yours alone. They sit in your settings, and OpenAI says ChatGPT considers them in every conversation from then on. They’re the right place for a personal voice and the wrong place for a company one, because nobody else can see them.
A project holds instructions and files for one kind of work, and on the Business, Enterprise and Edu plans you can share it with your workspace. Chat access lets colleagues use the project’s instructions and files; edit access lets them change them. Shared projects are off by default on Enterprise and Edu until an admin turns them on.
A custom GPT is a packaged assistant with its own instructions and up to 20 knowledge files, each up to 512MB. In a workspace you can share it with named people, with anyone who has the link, or publish it to the whole workspace. The two permission levels are “can chat” and “can view configuration”; OpenAI’s page doesn’t describe an edit role, so the person who built it owns the changes.
What to put in
Start with a voice description of a few hundred words, written as instructions rather than adjectives. “Confident, approachable and professional” tells the model nothing it can act on. Telling it to open with the point, use contractions, give the figure rather than “significant”, keep one idea per paragraph and never use “solutions” gives it something to follow and gives you something to check against.
Then add the examples. Three or four pieces of your own writing that you’d be happy to see copied, in the formats you use most. A proposal introduction, a client email, a LinkedIn post, a page from your site. In our experience the examples do more work than the description, because the model picks up rhythm and word choice from them that nobody manages to describe.
Add your facts separately and keep them short: what you do, who for, the figures you can stand behind, the clients you’re allowed to name. Date the file, because you’ll need to know later whether it’s current.
Finish with a list of words and phrases you never use. Every company has them, and the model will reach for them unless told.
Where it stops working
The first place is checking. A project tells ChatGPT what you’d like. Nothing reads the draft back against the instructions and tells you where it slipped, so the setup that worked in March has stopped matching the drafts by June and nobody’s noticed. You check by hand or not at all.
The second is the team. Custom instructions are personal, so ten people on personal plans have ten versions of the voice. Sharing a project or a GPT fixes that on a business plan, with one owner editing it and no record of what changed when.
The third is facts. A number pasted into a project in January is still there in August after the price has changed, and the model will use it with complete confidence. OpenAI’s knowledge pages describe how files are retrieved; they don’t describe any check of what the model says against what the files contain.
What people get wrong
They write the description as a list of adjectives, then wonder why every draft sounds like every other brand that chose “friendly and professional”. They skip the examples because the description feels thorough. And they treat the setup as finished, when the most useful thing about a voice description is that each time a draft comes out wrong, you can work out which line it broke and sharpen that line.
How BeigeCheck does it
BeigeCheck holds your company’s approved voice, glossary, house rules and facts in one place, and ChatGPT reads them through a read-only connection whenever you name BeigeCheck in what you ask. The same knowledge reaches Claude, and Microsoft 365 Copilot on Business and above, so nobody rebuilds the setup when they switch tools.
Each person also has their own voice, built only from writing they’ve done themselves, so a colleague keeps their voice inside the company’s standards.
After ChatGPT writes, the Beige-ometer checks the draft: a score for how authentic it reads and one for how well it fits the voice, the sentences worth changing with a fix for each, and every claim checked against the facts you’ve approved. When people keep making the same edit, BeigeCheck proposes a change to the voice, and a person approves it before anything moves.