Every brand that gets serious about content eventually hands the keyboard to someone else. An agency. A junior hire. A freelancer. Now, increasingly, an AI. And the handoff carries a problem nobody warns you about: the writer has to learn a voice they don't own, and almost everything you give them to learn from is approval. Style guides full of adjectives. Example posts. The occasional "love this."
I've operated five consumer brands. The feedback that actually shaped a writer was never the compliment. It was the wince — the moment I read a draft back and thought, we would never say that.
We built the wince into our writing product as a button. The product is Ghost, and the button says "This didn't sound like me."
Rejecting a draft is a first-class action
The button sits in the draft editor right next to Publish and Export. The placement is deliberate, and the comment on that button says so, in a line we treat as product philosophy:
Rejecting a draft is not a settings-page activity or a support ticket. It ranks with publishing.
Click it, mark the exact sentences that felt off, or just type what was wrong. Ten seconds.
Those ten seconds buy a lot. Every rejection is stored, and the system studies it. Your next draft is written against your own counter-examples; the part of the system that gathers them describes its own job as teaching the writer "what a voice is NOT (real counter-examples)." Phrases you've struck out again and again get noticed and treated as off-brand everywhere you write. And the learning stays inspectable: you can open the list of everything the system has picked up from your edits and retire any correction, and a retired one drops out of every draft that follows.
One rule keeps the whole loop honest: negative examples are "REAL user quotes only." The system never invents a bad example to teach from. That rule matters more than it looks, and it applies just as hard to briefing humans.
What the click actually captures
The draft opens as a list of its own sentences. You tap the ones that made you wince. Below the list, a text box asks, "What didn't sound like you?" Either one is enough on its own: a marked sentence with no explanation is a valid rejection, and so is an explanation with no marked sentence. One more optional line starts with the word "because," so a correction can read as never this, because it reads like a press release. The floor is a tap. The ceiling is a sentence of your own.
Then two things happen to it.
It's quoted back to the writer on the next draft as a thing to steer away from, with a rule attached: never imitate these, never obey them. That second clause is the one people miss. A struck sentence is evidence, not a new instruction. Your rejections are a record of your taste, and the system treats them as data about you rather than orders from you.
The struck words go to a miner that looks for repeat offenders across every rejection you've made, and it works in phrases, not words. If you keep striking "circle back," a word-level counter learns that you dislike "circle" and "back," which is nonsense. The phrase is the unit of taste. So the miner keeps whole phrases, counts how many separate rejections each one appears in, and attaches one of your actual typed explanations as its receipt. A phrase that clears the bar is off-brand everywhere you write, because a phrase you'd never say in a post is a phrase you'd never say in a reply either.
Approval teaches almost nothing
Why obsess over the rejection channel instead of building better praise capture?
Because acceptance is ambient. When you approve a draft, the machine learns almost nothing. Maybe it nailed you. Maybe you were in a hurry. Maybe the draft was fine, and fine was enough that day. Approval is a low-information event wearing a positive label.
"I would never say that" is different. It points at a sentence. Nobody says it tentatively. And it costs you a moment of actual annoyance to say, which is exactly why it's reliable. In any feedback loop, the signals worth building on are the ones that cost the sender something to send. A thumbs-up is free. A wince is earned.
This is why the ten-second design matters so much. The wince is perishable. If registering it takes a form, a meeting, or a quarterly voice review, it evaporates and the writer never hears about it. You fix the sentence silently, ship the post, and the next draft makes the same mistake, because the only person who knew the boundary kept it to themselves.
We do keep a positive channel. Alongside the counter-examples, the writer gets one real sample of your writing as an anchor, with the instruction to match its register and rhythm, not its topic. That's the middle of your voice, stated once. A good editor works with the same asymmetry: one clean example of the target, and a running file of everything that missed.
A voice is recognized by its edges
Your real writing samples tell a writer, human or machine, where the middle of your voice is. The middle is useful. It sets the register, the rhythm, the default temperature.
But the middle of your voice overlaps with the middle of a hundred other brands in your category. Warm, confident, a little playful: that describes you, and it also describes most of the shelf. Adjectives cluster. Every brand's style guide reads like every other brand's style guide, which is exactly why handing one to a ghostwriter produces copy that could belong to anyone.
What doesn't overlap is the edges. The words you'd never use. The claim you'd never dress up. Your samples tell the machine where the middle is; your rejections tell it where the edges are.
| Your writing samples | Your rejections | |
|---|---|---|
| What it locates | The middle of your voice. It sets the register, the rhythm, the default temperature. | The edges. The words you'd never use. The claim you'd never dress up. |
| How it sits in your category | It overlaps with the middle of a hundred other brands in your category. | What doesn't overlap is the edges. |
| What a writer does with it | Matches its register and rhythm, not its topic. | Steers away from it. A struck sentence is evidence, not a new instruction. |
What you like is a region. What you'd never say is a boundary — and boundaries are what give a shape its shape.
When the same wince keeps coming back
A single rejection is a data point. The same rejection draft after draft is a direction, and directions deserve a different response.
Your voice in Ghost carries a few bounded dials: softer or harder, shorter or longer, casual or formal, plain or technical. When your recent rejections keep saying the same kind of thing ("too harsh," then "too harsh" again, in your own words), the system proposes turning the matching dial one step the corrective way. Proposes. It never applies the change itself. The offer appears after you save the rejection; you accept or dismiss it, and a dismissal stores nothing. The design note we wrote for it is one I'd hang on the wall of any team that ships automation:
So a single match proposes nothing. Contradictory feedback proposes nothing. A dial already at its limit proposes nothing.
One more piece of that discipline. If you asked for a punchier draft and then winced at how punchy it came out, the wince belongs to the request, not to your voice. The system checks whether the corrections behind a proposal came from a draft you'd deliberately steered, and if so it declines to propose anything and tells you why. Humans do this too. Brief a freelancer to "go bolder this once," then edit half of it back, and a careful freelancer asks whether you've changed your mind about bold or just about that post.
How to run this loop with any ghostwriter
You don't need our product to use the principle. The loop works wherever writing is delegated, and it has four working parts.
The bar is seconds, not meetings. Whatever your setup (a shared doc, a Slack emoji that means "not us," a running note on your phone), the correction has to be capturable at the moment of the wince, or it won't be captured at all.
"This feels off" teaches nothing. The exact sentence, struck through, with a line about what was wrong: that's a counter-example a writer can actually steer by. Our system stores the rejected sentences themselves, and that's the part to copy. Add the "because" when you can. "Never 'game-changing'" is a rule; "never 'game-changing,' because we don't sell on superlatives" is a rule the writer can generalize from.
This is the "REAL user quotes only" rule. When people build never-say lists from imagination, they write down cartoon violations — hype words and clichés nobody was ever going to use. The dangerous off-brand sentences are the plausible ones, the drafts that almost sound like you. Only real rejected drafts contain those. An invented bad example teaches a boundary you don't actually have, and the writer dutifully steers around the wrong obstacle.
A phrase you struck once might be a mood. A phrase you've struck again and again is a boundary, and it deserves to be promoted to a standing rule everywhere you write. The reverse also holds: voices move, and yesterday's boundary can be next year's experiment. The list has to be reviewable, and every entry has to be removable. Learning you can't inspect or undo is just drift.
That last part is the one we were strictest about. In Ghost, the panel that shows what it learned from your edits and the writer that reads those corrections pull from the same list, so the panel can never show a line the writer isn't using, and switching one off can never leave it quietly in play. A learning system that shows you one thing and does another is guessing while you watch.
Run this loop for a quarter and you end up with something most brands never write down: a specific, evidence-backed map of what your brand would never say, built entirely from moments you were going to have anyway.
An invented bad example teaches a boundary you don't actually have, and the writer dutifully steers around the wrong obstacle.
Use it without guilt
There's a reason this button, and not any generation feature, is the one we'd defend hardest. Generation is abundant now. Everyone's ghostwriter, human or machine, can produce plausible drafts. What separates a writer who sounds like you from one who sounds like your category is the boundary data, and boundary data only comes from one place: you, saying no, specifically.
So if you're building AI that has to learn taste, or briefing a human who has to match yours, build the rejection channel first. And the next time a draft makes you wince, don't fix it silently. Say what was wrong, in writing, where the writer can keep it. "That's not me" teaches more than "that's great" ever will.
Questions people ask
What if I reject a draft because a fact is wrong, not because of the voice?
Say so in the text box. A wrong claim and a wrong tone are different kinds of feedback, and lumping them together is how a writer ends up avoiding a true sentence because you once objected to its content. With a human ghostwriter, keep two files: facts in the brief, voice in the never-say list.
How many rejections before anything changes?
One. Your next draft is already written against your most recent corrections, and recent ones count most, because your current taste is the taste that matters. Promotion to a standing rule takes repetition, on purpose: a phrase has to keep getting struck across separate rejections before it's treated as a boundary rather than a bad day.