How to Make ChatGPT Sound Like Your Brand (Not Everyone Else’s)

There’s a specific silence that happens when you paste AI copy into the CMS and read it one last time before you hit publish.

It’s fine. You know it’s fine. You also know it’s nothing. A beginning, a middle, three benefits, and a call to action that could sell renters insurance. Nobody’s going to complain about it. Nobody’s going to remember it either.

You publish it anyway, because it’s 5:20 and fine was on the calendar and good wasn’t.

Then comes the diagnosis everyone lands on: the prompt wasn’t detailed enough. Next time, more instructions.

That’s the wrong conclusion, and it’s wrong in a way that quietly costs you the next two years of content.

ChatGPT sounds generic because you handed it an assignment instead of a point of view. Three things fix that and none of them is a better prompt. Custom instructions that persist across every chat. A body of your own writing it can study before it drafts. Corrections written as rules instead of line edits.

The short answer

• Custom instructions. Standing rules, written once, applied to every chat. Use them for anything with a yes-or-no answer. Banned words, structure, formatting. Never adjectives.
• Context priming. Paste three to five of your best pieces and make the model tell you the pattern back before it writes a word. Almost nobody does this step. It’s the one that moves the most.
• Voice calibration. Correct in rules, not line edits, so each fix holds for the rest of the session.
• Try this first: open a chat, paste your three strongest pieces, and ask the model to describe your voice back to you. Its answer tells you exactly how far you have to go, and occasionally it tells you something worse. More on that below.

Bottom line: you can’t prompt your way to a voice. You can only hand over enough evidence that averaging stops being the model’s best guess.

Every Model Writes Toward the Middle. That Is the Entire Problem.

Ask for a blog intro about project management and the model averages everything it has ever read about project management, then hands you the exact center of that pile. The center is safe. The center is also where your whole category is standing, because everyone asked the same averaged-out question.

Your voice is not in the average. Your voice is the pile of small deviations from it. The phrase you reach for every single time. The argument you’ve made at four conferences and will make at a fifth. The joke your CEO tells in every webinar whether or not it lands. The claim you refuse to make because it sounds like a pitch.

None of that is in the middle, and a bare prompt gives the model no reason to leave the middle.

So it doesn’t. You get “in today’s fast-paced business environment.” You get three benefits and a tidy conclusion. It’s not wrong. It’s just nobody. And nobody is worse than wrong, because wrong at least gets noticed.

A Police Sketch Is What Adjectives Get You

Ask a witness what the guy looked like and you’ll hear average build, brown hair, normal face, maybe thirties. The sketch that comes back is technically a face. It resembles nobody in particular. You could hold it next to four hundred men and get a maybe on most of them.

The artist isn’t bad. The artist got adjectives.

“Write in a bold, conversational, authoritative tone” is average build, brown hair, normal face. It describes four thousand companies. The model has no way to know that your bold is dry and understated while the agency across town means all caps and three exclamation points. Descriptors point at a neighborhood. They don’t give an address.

There’s a whole economy built on selling you a longer, cleverer prompt, and it will never say this out loud, because “build the system once and stop buying prompts” is a hard product to move at $49 a pack.

Two things beat a longer prompt, and neither of them is a prompt.

Persistence. The rules live somewhere the model reads every time, so you’re not retyping context from memory and hoping you remembered all of it. A prompt is a single transaction. Every new chat drops you back at the middle.

Evidence. You show the model your voice instead of describing it. This is the police sketch problem in reverse. Stop giving it adjectives. Give it the photograph.

The move	What it is	What it fixes	Sticks around?
Custom instructions	Standing rules in ChatGPT’s settings, applied to every chat	The obvious tells: banned words, formatting, structure	Yes, until you change them
Context priming	Your best work, pasted in, analyzed before it drafts	The texture no rule can describe: rhythm, how you open, what you refuse to say	Per chat, unless you save it to a project
Voice calibration	Corrections written as rules, not line edits	The last 15%, where the draft is close and still not yours	Within the session, permanently if you fold the rules back into your instructions

The Second Cost Is the One People Miss

The first cost of anonymous copy is obvious. A reader who can’t tell your post from four others doesn’t remember who wrote it, so you paid to build your category’s awareness instead of your own. Eventually you hear about it from the VP who forwards you a competitor’s blog post at 11:40 at night with the subject line “why don’t we sound like this.”

The second cost is newer and it compounds.

When a buyer asks ChatGPT or Perplexity “what’s the best project management tool for a hospital marketing team,” the engine builds its answer out of sources that hold a clear, repeatable position. Bland pages get averaged back into the background they came from. Specific pages get pulled out and named. The flatness that makes your blog forgettable to a person is the same flatness that makes it invisible to the system deciding who gets recommended, which is most of the mechanic behind how organizations show up in AI search results at all.

Generic isn’t a style problem. It’s a visibility problem wearing a style problem’s clothes.

Custom Instructions Are Rules, Not Vibes

Custom instructions are standing rules you write once. The model applies them to every conversation until you change them. Two boxes: what it should know about you, how it should respond. They live under Settings, then Personalization, then Custom Instructions.

Most people leave them empty or type “be concise” and call it handled. Concise is table stakes.

Steal this structure and swap in your own:

HOW TO RESPOND

Structure
- Lead with the reader's problem. No product until the problem lands.
- One idea per paragraph. If a sentence survives being cut in half, cut it.
- No "Related reading" lists. Work the reference into the sentence.

Diction
- Never use em dashes. Restructure instead.
- Banned: leverage, streamline (with no object), robust, best-in-class,
  seamless, "in today's landscape."
- Contractions are fine. Write "you" and "your team," never "users."
- One exclamation point per piece. Ideally zero.

Evidence
- Never invent a statistic, a customer name, or a product capability.
  If you need one and don't have it, write [NEEDS SOURCE] and keep going.

None of those are descriptors. Every one has a pass-fail condition. “Be bold” is a vibe. “Ban these six words” is a policy a model can follow, and you can check whether it did in about nine seconds with a text search.

That last block does more work than it looks like. A model told to sound confident will invent a number to sound confident with, and it will do it in the same tone it uses for things it knows. Giving it permission to say “I don’t have this” is the difference between a draft you fact-check and a draft you rebuild. The same instinct governs what goes into the prompt, which stops being a style question the second you’re working anywhere near patient data.

Custom instructions are the floor. They stop the obviously off-brand things. They won’t teach texture.

Stop Describing Your Voice. Hand Over the Evidence.

One move separates the people getting usable drafts from the people still fighting the tool. Before you ask for a single word of new copy, you give the model a body of your existing writing.

That’s context priming. Paste three to five of your best pieces (the ones you’d frame, not the ones that hit quota) and tell it to study them before it writes anything. Not “write like this.” Study this. You’re asking it to pull the pattern, not trace the surface.

Below are four examples of our best content. Before you draft anything,
analyze them and tell me back what you notice: sentence length and rhythm,
how we open, how we handle transitions, what we consistently do and
consistently refuse to do, and the texture of the voice. Be specific.
Then wait for my brief.

Two things happen. The model builds a picture of your voice out of evidence instead of adjectives. And you get to see whether it found you, which is a free diagnostic almost nobody runs.

So grade it.

“Your voice is professional, informative, and engaging” is a horoscope. Translation: it found nothing and it’s covering.

“You open with a specific failure the reader has lived. You never hedge with ‘can help.’ You quote the customer instead of paraphrasing. You end sections on a short declarative line.” That one found you, and everything downstream inherits it.

Now the uncomfortable part. Sometimes you feed it your genuinely best work and it still hands back the horoscope. The reflex is to blame the model. The more likely reading is that there’s no pattern in the samples to find. A model priming on twenty thousand words of your own writing is a mirror with a delay, and if it can’t locate a voice in there, the problem was never the tooling.

If that’s where you land, swap in the pieces that made someone internally a little uncomfortable before they went out. More signal in those. And if you don’t have any of those, you’ve just learned something more useful than a prompt.

You can also prime on what to avoid. Paste a slab of thoroughly generic competitor copy and say: this is the trap, never sound like this. One good negative example beats ten positive ones, because it names the exact ditch you keep steering into.

Prime the positioning too, not just the prose. Who you’re for. Who you’re not for. The opinion nobody else in your category will say out loud. The proof you’ve earned the right to claim. Voice without a point of view is just a nice-sounding average, which is why the vertical prompt sets we publish for healthcare marketing and higher-ed marketing teams open with audience and stakes instead of output format.

Correct It Like an Editor, Not Like a Client

Even after instructions and priming, draft one lands close and not yours. This is where most people quit, and they quit in one of two directions. They rewrite the whole thing by hand, which wastes the tool. Or they ship the almost-right version, which is how off-brand copy gets out the door with nobody’s fingerprints on it.

Calibration is the third option. Correct it the way an editor corrects a promising junior writer: every note teaches a rule, not a line.

Weak correction	Strong correction	Rule it installs
“Make this better.”	“This paragraph hedges. We don’t hedge. State the claim, and from now on flag any sentence softening with ‘can help,’ ‘might,’ or ‘often.'”	No hedging verbs
“The intro is boring.”	“That intro describes a problem in the abstract. Open on a moment the reader has lived, with a person and an object in it. Do that for every intro.”	Concrete openings
“This doesn’t sound like us.”	“You wrote ‘organizations can leverage.’ We never name an abstract actor. Say who does the thing: your ops lead, the designer, whoever owns the calendar.”	Named actors, not abstractions

Four or five of those across a draft and something clicks. The model stops making that entire class of mistake and the next draft lands closer on the first pass. You’re not editing a piece. You’re training a session.

It will also overcorrect, and nobody warns you about that either. Teach it that you open with a lived moment and it opens every section with a lived moment. Teach it that you write short and everything turns staccato until the whole thing reads like someone having a small crisis in a parking garage. It picked up your moves without picking up how often you use them, because frequency is not something five samples can teach.

Two fixes. Prime on range, which means putting a long analytical piece in next to your punchiest one. And name the once-per-piece moves out loud: the specific-moment opening goes at the top of the article, not the top of every section.

When a session finally sings, save it. Ask the model to summarize the voice rules it learned. Read the summary, because it’ll get one or two wrong, confidently. Fold the accurate ones into your custom instructions. Otherwise every bit of that work evaporates the second you close the tab.

The Same Brief, Ten Minutes Apart

One variable between these. Whether the model had been primed.

Before:

In today’s fast-paced marketing environment, managing multiple projects across teams can be challenging. Organizations often struggle with visibility and alignment. A centralized approach helps teams streamline their workflows, improve collaboration, and drive better outcomes.

After:

Your designer just asked which version of the banner is final. Fourth time this week somebody’s asked you a question a shared file could’ve answered. At that point you’re not managing projects. You’re managing the distance between where the answer lives and where people go looking for it.

What changed isn’t tone.

The first one has no person in it. No object. No time. Its subject is “organizations,” which is nobody. Every verb is a category verb: manage, struggle, streamline, drive. Paste it under nine hundred logos and it fits all of them, which is the same thing as fitting none.

The second one has a designer, a banner, a count, and a claim you could argue with. Nothing about it is cleverer. It’s more specific, and specificity is most of what people mean when they say something sounds like an actual company wrote it.

One rule out of this whole article, take this one: make it name something.

The Whole Thing, Start to Finish

You don’t need any of this for a Slack message or a routine follow-up, where a decent prompt genuinely is enough and a set of ready-made Claude email prompts gets you there in a minute. You need it for anything with your name on it in public.

1. Set custom instructions once. Hard rules, banned words, permission to flag what it doesn’t know.
2. Prime before you brief. Best three to five pieces plus your positioning. Make it analyze. Grade the analysis.
3. Brief specifically. Topic, audience, and the one job this piece has to do.
4. Calibrate in rules. Four or five corrections, each one teaching something.
5. Save the pattern. Summary, checked, folded back into your instructions.

First pass costs an afternoon. Every pass after costs minutes, which is most of the reason a one-person marketing function can now ship like a three-person one, and the whole premise behind AI prompts for a marketing team of one.

The Prompt Was Never the Hard Part

“Here’s a prompt you can steal” keeps disappointing people because a prompt is a single transaction. It does one job and forgets you the moment the chat closes. What makes AI writing sound like a specific company is a small system that remembers who you are. Rules that persist. Examples that teach. Corrections that stack.

Set it up once and the tool stops fighting you. Drafts land closer. Edits get shorter. Your copy stops reading like it came from the same averaged-out place everyone else is pulling from.

Your voice already exists, or it doesn’t.

If it does, it’s sitting in a folder nobody has opened since it published, and the machine you keep asking to imitate it has never once been shown it.

If it doesn’t, no prompt was ever going to cover for that.

Want to know where your team actually stands with AI? Take the AI readiness scorecard. Five minutes, no sales call.

Frequently asked questions

Why does AI writing sound generic? Because a language model writes toward the average of everything it’s read on a topic, and a bare prompt gives it no reason to deviate. Your voice is a set of specific choices that live outside the average, so unless you feed the model those choices through custom instructions, examples of your writing, and iterative correction, it defaults to the safe, anonymous middle every other brand lands in too.

How do I give ChatGPT my brand voice? Three things a prompt alone skips. Set custom instructions with your hard rules and banned words so they apply to every chat. Prime the model on three to five of your best pieces and make it analyze the pattern before it writes. Then calibrate its drafts by correcting in rules (“we never hedge”) rather than one-off line edits, so it learns the pattern instead of fixing one sentence.

What are custom instructions in ChatGPT? Standing rules you write once that ChatGPT applies to every conversation. They live under Settings, then Personalization, then Custom Instructions, and there are two fields: what the model should know about you, and how it should respond. For brand voice, use them for rules with a clear pass-fail condition such as banned words, structural rules, and formatting bans, not vague descriptors like “be professional.”

Does ChatGPT remember my brand voice between chats? Only what you’ve stored. Custom instructions and saved projects carry forward. Everything you teach it inside a single conversation, including all your calibration corrections, disappears when that chat closes. That’s why the last step matters: ask the model to summarize the voice rules it learned, check the summary, and move the accurate ones into your custom instructions so the work compounds.

Isn’t a very detailed prompt enough to fix generic output? It helps, then it hits a ceiling. Adjectives like “bold” or “witty” describe thousands of brands and don’t tell the model your version. And a prompt is disposable, so you’re retyping context every time and hoping you remembered it all. Persistent instructions plus examples of your actual writing beat even a very long prompt, because they give the model evidence instead of descriptions.

How long does it take to set this up? The first full pass, meaning writing custom instructions, gathering your best examples, priming, and calibrating a draft, takes an afternoon. After that each piece takes minutes, because the setup carries forward and you’re only briefing and lightly correcting. The setup cost is one-time. The payoff repeats on every piece you produce.

---