Chat That Knows You're the Author, Not the Audience

If you run a brand and you've asked an AI chat tool to write for you, you've probably noticed the ceiling. The drafts are competent. They're fast. And they sound like a smart stranger doing an impression of a company roughly like yours.

That's not a quality problem you can fix by switching to a smarter model. It's a framing problem built into how these tools see you, and understanding it will change how you use every AI writing tool you touch, including ours.

A generic assistant treats you as the audience. You ask, it performs, and what it performs is tuned to the average asker, the statistical center of everyone who ever typed a question like yours. For research or code, that's fine. For writing in your brand's voice, the average asker is precisely the thing you're trying not to sound like.

The fix is a different relationship. The tool has to know you're the author.

Before it writes a word, it reads a briefing

This is what it looks like in practice, in Ghost, our writing product. Before the chat writes a word for you, it reads a briefing about you.

Not your topics. Who you are on the page.

One observation per category, highest confidence first. A briefing, not a transcript dump.

That last line is a design decision, and the temptation runs the other way. Once a system starts accumulating observations about a writer, the obvious move is to shovel all of it into every prompt. More context, better output: that's the intuition. It's wrong the same way it's wrong for humans. Hand a freelancer forty pages of brand documentation and they'll write from none of it. Hand them one page of the highest-confidence, most current facts about your voice, and it shows up in the second paragraph. Curation is what makes context usable, for machines exactly as for people.

If the word "briefing" sounds abstract, the voice profile itself is not. On Ghost's features page we show three example records, none of them a customer, each built the same way and each reading like a different person. Every record has the same five fields: how you sound, your sentence rhythm, the words you reach for, the words you'd never use, and how you close. One example record reads "Deadpan. Confrontational. Anti-corporate." for how it sounds and "Short. Declarative. ALL CAPS for emphasis." for rhythm, with a never-use list of wellness, journey, thrive, optimize. Another reads "Considered. Literary. Restrained.", closes "Quietly. No imperative. No urgency.", and bans hack, game-changer, drop, slay. Same machine, same five fields, and you would never mistake one for the other. That is what the chat reads before it writes. Not a mood board. A record with fields, and the fields are the ones that decide whether a paragraph sounds like you.

One instruction, run twice

The cleanest way I know to show what the briefing changes is a demo we put on Ghost's own page. One instruction, run twice. The brief was a single sentence: "Write one line introducing the protein." The brand was a paleo protein brand we use as an example, not a customer, and the two lines came back like this.

The first line is what any model writes by default, and that is the point. It isn't bad. It is the statistical center, the sentence every protein brand's website already carries, and it could belong to anyone. The second line names a real fact about the brand, the four-ingredient wedge, and follows a slang policy the brand actually has. Nothing about the second line required a smarter model. It required the model to know two things that live nowhere on the public internet, and the record is how it knew them. We let readers of that page pick which line they think is Ghost's, and the pick is the whole demonstration. If you can tell, the record is doing work a prompt can't.

Then it grades its own impersonation

Writing as someone raises an honesty question: how do you know the impersonation held?

So after every draft it writes, the system grades its own work against the same profile it wrote from: matched phrases, identity signals, evidence quotes pulled from the draft itself. Not a vibe score. Receipts: this is where the draft used your phrasing, this is what marks it as yours.

The same habit runs on every post before it can publish, not only in chat. A draft comes back with a verdict per check and a finding in plain words. On the features page we show what those findings look like. A voice check that passes reads "Tone deadpan, ALL CAPS opener, sentence length on-spec." A voice check that drifts reads "Voice score dropped to 72/100 - too corporate. Cut the wellness framing, return to deadpan." A brand check reads "No banned words," then names the pillars the draft hit. Every one of those is a sentence a marketer can act on, because it says what the draft did and against which line of the record it was judged.

If you take one evaluation habit from this piece, take that distinction. Any measuring tool (a voice checker, an analytics dashboard, an audit) will sometimes fail to measure. The honest ones tell you "the check didn't run." The dangerous ones render the same red badge for "this is bad" and "we don't know," and let you act on the difference as if it weren't one.

A bigger model can't close the gap

This design matters more than which AI model sits underneath, and the reason is specific.

Your banned list isn't in anything the big models were trained on. Neither is the observation about how your cadence actually behaves, or the phrase your founder uses that no competitor would.

The only way a model knows any of it is that you told it, and telling it is a system, not a one-off request.

That's the asymmetry hiding under all the model-race noise. Every brand in your category can rent the same intelligence, from the same vendors, at the same price. Capability is a commodity with a subscription page. What can't be rented is the accumulated, curated, current record of who you are on the page — because you're the only one who has the raw material. It isn't writing for you. It's writing as you.

What if it learns the wrong thing about you

The objection I hear most, once people understand that the system accumulates observations, is the obvious one. What if it learns something wrong? A week of off-voice drafts, a guest writer, a campaign where you deliberately sounded unlike yourself. Does that get filed as fact about you?

Two things stop it. The first is the ranking inside the cap: within each category only the highest-confidence observation rides into the briefing, so a shaky note about one strange week doesn't get equal billing with a pattern the system is sure of. The second is that the record isn't hidden. You can read the voice record Ghost builds, edit the identity it wrote, and switch off any learned line you disagree with. That is a deliberate position, not a convenience. The tool's model of you is a document you own, not a black box you are subject to.

The briefing is curated by the system. The last editor of it is you.

You can build the paper version today

You don't need our product to act on this. The briefing translates directly into a document you can maintain by hand and paste into whatever AI tool you already use. Ours has three sections; yours should too.

The models will keep getting better, for you and for every brand you compete with, simultaneously. None of that movement changes your relative position. The briefing does.