While we were building Ghost, our writing tool, we ran one instruction twice. "Write one line introducing the protein." The brand was Paleo Pro, one of the example brands we use to show the product. It isn't a customer; it's owned by my manufacturing partner, so it's a brand I actually know. The first run came back: "Premium nutrition for an active lifestyle." The second: "Four ingredients per scoop. All cow, no bull."
| Any model, by default | Written from the brand | |
|---|---|---|
| "Write one line introducing the protein." (EXAMPLE · Paleo Pro · not a customer) | "Premium nutrition for an active lifestyle." | "Four ingredients per scoop. All cow, no bull." |
The first line is what any model writes by default, and it could belong to any protein brand on any shelf. The second was written from the brand. It names a real fact, the four-ingredient wedge, and it uses the brand's own slang policy, which allows "all cow, no bull" and would never allow "active lifestyle." Same instruction, same model. The difference is everything the second run knew before it started.
You can run the same test on your own brand this afternoon. Ask any model for one line introducing your product, then check whether the line could sit on a competitor's shelf without a word changed. If it could, that is the flattening, and you've just measured it.
Your brand has a voice. AI keeps flattening it. With nothing of yours to read from, a model writes the line that would work for any brand on the shelf. Yours isn't supposed to. Ghost is the product we built to write from your brand, and it is slower than a tool that takes a prompt and posts, because nothing goes out until you approve it. This piece is about what it holds, how it uses that during a working week, and what a small team stops paying for once the voice lives somewhere other than one person's taste.
What the record holds
Before Ghost writes a word, it assembles a voice record. It builds that record by interviewing you about how your brand sounds, reading what you've already published, and taking the modifiers you name. It doesn't scrape your old posts and imitate them.
The record reads like the description a good editor would write of you. Take one of the example records we built to show the product, for a drinks brand that wants nothing to do with wellness copy. How it sounds: "Deadpan. Confrontational. Anti-corporate." The words it reaches for: "murder, death, evil, kill, certified, infinite." The words it would never use: "wellness, journey, thrive, optimize." How it closes: "Imperative verb. No CTA softener." You can read your own record, edit the identity it wrote, and switch off any learned line you disagree with.
Voice tells Ghost how to write. Knowledge tells Ghost what to write, and what not to. The knowledge base holds the facts only you can cite: your numbers, your studies, your customer language, the founder story, the tagline approved for pack copy. Each entry carries its source. Approved once, cited every time, and the line gets used right or it doesn't get used at all. Generic AI has neither the voice record nor the knowledge base, which is why it wrote "premium nutrition for an active lifestyle."
Then modifiers. Same voice, different volume: named, stackable adjustments for a channel, a style, a tone, a format, or anything you write yourself. Tighter. Twenty percent more dry humor. For LinkedIn. A modifier moves the draft around inside the voice without leaving it, which is what a good writer does when you say "same thing, but for the newsletter."
The voice record is one part of the larger brand record every Jinn product reads from, 346 signals of what your brand is, and that record has a piece of its own. Ghost is the product that writes from it.
Voice tells Ghost how to write. Knowledge tells Ghost what to write, and what not to.
Six verbs, one week
Ghost runs your whole content process, and the process has six verbs: research, plan, write, schedule, monitor, configure.
A working day looks like this. Overnight it read your category, the publications, communities, and competitor moves that matter for your brand, watched and ranked. In the morning it hands you a brief: hook, angle, supporting points, already drafted from what's happening and what you're trying to say. The strongest signals arrive with a draft attached; you decide whether to send it. The draft waits in the queue, in your voice. You read it, tweak a line, approve it or send it back. It publishes on the cadence you set, to X, LinkedIn, Bluesky, and Facebook. On Monday it grades last week's posts on X and LinkedIn, the channels it reads performance from, so you know which ones actually worked and the next brief starts smarter than the last.
It read your category, the publications, communities, and competitor moves that matter for your brand, watched and ranked.
Hook, angle, supporting points, already drafted from what is happening and what you are trying to say.
The draft waits in the queue, in your voice.
You read it, tweak a line, approve it or send it back. Nothing goes out until you approve it.
It publishes on the cadence you set, to X, LinkedIn, Bluesky, and Facebook.
It grades last week's posts on X and LinkedIn, the channels it reads performance from, so you know which ones actually worked and the next brief starts smarter than the last.
The writing itself has three doors. You have a topic: give it the topic and get a post in your voice, in the right format for the channel. Something happened: drop in the post, the article, the competitor move, and get three takes in your voice, agree, disagree, reframe, however you want to enter the conversation. Or you have a draft that's fine and doesn't sound like you: paste it, and what comes back does, with names, numbers, and personal details guarded so the rewrite can't quietly change a fact. Long-form has its own room, the atomizer splits one long piece into many reviewed posts, and the hook bank keeps the openers that landed.
The hardest copy, and the review chain
The hardest writing I ever managed for a brand was aimed at people who had already bought. After the purchase the tone has to change. You're nurturing, educating, occasionally selling again, and every message has to stay on-brand without going generic, to a customer who already knows what the brand sounds like and will notice the moment it doesn't. Even with a really good team it takes a lot of turns before you can trust them to execute that without you in the room, giving feedback on every piece and reading each one critically.
Two things in Ghost are built for exactly that kind of trust. The first is the record: the voice and the knowledge are the same for the message to a repeat customer as for the launch post, so the fourth writer on the fourth channel starts from the same page as the first. The second is the review chain, which has four seats. The Owner runs the brand and signs anything. Under that sits a Brand Editor who owns the standard, the person who would have caught "active lifestyle" before it shipped, and who approves anything that goes out; Brand Writers, who draft and hand up the chain; and a Brand Viewer seat for the stakeholder who needs to see the work and shouldn't be able to touch it. Work moves up the chain, and a writer never approves their own post. For an agency, every client brand gets its own record, its own roles, and its own review chain, in one workflow.
They draft and hand up the chain.
Owns the standard, the person who would have caught "active lifestyle" before it shipped, and approves anything that goes out.
Runs the brand and signs anything.
When a draft is wrong in a way you can feel and can't articulate, there's a button for that, and the button got a piece to itself. The short version is that the wince teaches the record.
Three judges and a log that can't be edited quietly
Voice checks run on every post. Wherever you switch compliance on, three judges read the draft in parallel, and each returns a verdict with the exact finding. The regulatory judge reads for claims the law governs in your category; "improves immunity" on a drink is the kind of line it stops. The brand judge reads for banned words and whether the post hits the pillars you set. The voice judge scores the draft against the record and says what drifted: too corporate, cut the wellness framing, return to deadpan.
The example verdicts on the page show the three states. All clear. Voice drift, where the regulatory and brand judges pass and the voice judge flags a 72 out of 100 and tells you why. And a hard fail, where a draft claims the product "improves immunity," an unverified health claim, and the regulatory judge blocks it from shipping regardless of how good the writing was. If you sell supplements, finance, healthcare, food, alcohol, anything a regulator reads, that third state is the one that matters most. I ran a prenatal supplement, and the judgment that draft needed used to live in one person's head; the founder piece for this product is about that.
| What it reads for | In the example verdicts (EXAMPLE · three demo states) | |
|---|---|---|
| The regulatory judge | Claims the law governs in your category. | Blocks a draft claiming the product "improves immunity," an unverified health claim, regardless of how good the writing was. |
| The brand judge | Banned words, and whether the post hits the pillars you set. | Passes. |
| The voice judge | How the draft scores against the record, and what drifted. | Flags a 72 out of 100: too corporate, cut the wellness framing, return to deadpan. |
Every verdict lands in an audit log built so it can't be edited quietly. For a regulated brand that is audit defense built into the writing flow: when someone asks what was checked before a post went out, the answer is a record you can produce.
A record can be checked against, by a person or by the thing writing the draft, and that is the difference between a voice that lives in your taste and one that survives the fourth writer, the fourth channel, and the day you're not in the room.
What it replaces
The category Ghost replaces is the retainer. A content agency or a steady freelancer is a real monthly cost, and the voice still has to be re-briefed every time the roster changes. Every new writer starts from your taste and your feedback and takes turns to get there. Ghost holds the voice in the record and drafts every day. The economics are supporting proof; sounding like you is the point.
For an agency the math runs the other way. Serving one more client usually means one more writer to brief, manage, and QA. With each client's voice held in its own record, the roster scales without the headcount scaling with it.
None of that replaces a good writer. The good ones are the taste, and the record learns from them. What the record replaces is the re-briefing, the part of every retainer that was quietly buying someone re-learning your brand.
The cost I named at the top is real. Every post waits for your approval, and the judges add a step. We take the slower path on purpose, because the voice we're holding is one you intend to keep.
What the record replaces is the re-briefing, the part of every retainer that was quietly buying someone re-learning your brand.
Your brand already has a voice. The guides you wrote for it were advice, and advice goes unread. A record can be checked against, by a person or by the thing writing the draft, and that is the difference between a voice that lives in your taste and one that survives the fourth writer, the fourth channel, and the day you're not in the room.
See how Ghost works
The voice record, the knowledge base, the modifiers, and the review chain, walked end to end.
How Ghost worksDoes Ghost imitate my old posts?
No. It builds the voice record by interviewing you about how your brand sounds, reading what you've already published, and taking the modifiers you name. It doesn't scrape your old posts and imitate them.
Where does Ghost publish, and what does it measure?
It publishes on the cadence you set, to X, LinkedIn, Bluesky, and Facebook. On Monday it grades last week's posts on X and LinkedIn, the channels it reads performance from.
Can a writer approve their own work?
No. Work moves up the chain, and a writer never approves their own post. For an agency, every client brand gets its own record, its own roles, and its own review chain, in one workflow.