This year I ghostwrote more than 2,000 articles. Real work, shipped, published. Some ran under named bylines in high-visibility placements. Some were SEO pieces designed to rank and convert. Some were reputation work, the kind of writing that quietly repairs what a bad press cycle broke. I also write a Substack and produce political content, including investigative exposés where I name real people who employ real lawyers.
That last category changes everything. When you write about litigious people and get a fact wrong, the story is no longer about them. It's about you.
Accuracy in political and investigative writing isn't a best practice. It's the minimum condition for continuing to exist. One unsupported claim, one wrong date attached to the wrong person, one misattributed quote, and you're not just embarrassed. You're in discovery. That pressure shaped every decision I made about how to build a writing process capable of handling volume.
The year in the trenches
Two thousand articles sounds like a lot. It is. But the number isn't really the point.
The point is what was inside it. On any given week I might ghost a CEO's LinkedIn post in the morning, spend the afternoon on an online-reputation piece that had to be careful about tone for legal reasons, and then open a political exposé that needed sourcing I could actually defend in front of a lawyer, all before the same Friday deadline. Four assignments some weeks. Four completely different registers. Four different relationships to facts, to persona, to reader, and to the specific kind of trouble a wrong word could cause.
Volume at that range doesn't simplify your work. It multiplies the variables, and the variables don't care that you're tired.
I wasn't writing 2,000 versions of the same thing. I was writing 2,000 different problems, and each one required me to know exactly what kind of problem it was before I wrote a single sentence.
What writing at scale actually requires
There's a myth that ghostwriting at volume means producing reliable output from a fixed template. Get the brand voice document, run the brief through the process, ship the piece. The appeal is obvious to anyone building a repeatable operation.
It's also wrong, and working at scale is what proved it to me.
A CEO's LinkedIn post lives or dies on sounding like a specific human being with specific opinions. An investigative piece has to sound authoritative and fair even when it isn't being kind. An SEO article needs to satisfy a query and hold a reader, two things that often pull in opposite directions. An online-reputation piece is written for an audience that is partly skeptical and knows it.
None of those problems are solved by a brand voice document. They're solved by understanding what each assignment actually is: what it needs to accomplish, who is speaking, who is reading, and what will make this particular reader trust this particular voice on this particular day. Template thinking can't hold that. It collapses everything into the content-marketing register, which is fine for content marketing and wrong for almost everything else.
Why existing AI writing tools failed my ghostwriting workflow
I tried the existing AI writing tools. More than a few. They weren't bad. They were built for a problem that wasn't mine.
They were built around brand voice and templates: define your tone, load your guidelines, generate your content. That's a real use case. Just not mine. I don't need a consistent brand voice across thousands of pieces. I need whichever voice this piece demands, held accurately, for the length of one assignment.
The deeper problem was structural. Most tools run a single model through a single pass: take the input, produce the output, done. For simple content, that works well enough. For anything where accuracy and voice both have to be right simultaneously, it doesn't. The model is trying to plan the piece, pull in relevant information, check its own reasoning, maintain a specific register, and write clean sentences, all at once, without stopping to verify anything.
That's not how good writing works. It's not how I work. I don't draft and fact-check at the same time. I don't build structure while polishing prose. Those are separate cognitive tasks, and the single-pass architecture was compressing them into one moment.
I'd open the output, read the first three paragraphs, and feel the specific flatness of something that had tried to do too many things at once and succeeded at none of them completely. Fluent. Competent. Indistinguishable from the ten other pieces on the same topic that week.
What a multi-agent ghostwriting tool actually does differently
So I built my own tooling.
The core idea was simple: stop asking one model to do everything at once. Break the job into its actual parts and give each part to a specialized agent. Planning is its own step. Research is its own step. Fact-checking is its own step. Editing is its own step. Each agent does one thing well and hands off to the next.
When the fact-checking agent's only job is to check facts, it checks facts.
It isn't also trying to maintain narrative momentum or hit a word count. When the editing agent's only job is to make the voice consistent and clean, that's all it's doing. The division of labor isn't organizational tidiness. It's what made the accuracy bar survivable and the voice bar possible, and it's the reason I could write a political exposé with the factual discipline that can withstand a lawyer's reading, then turn around the same afternoon and write a LinkedIn post that sounds like a real person rather than a content-marketing platform.
Building it took a while. Longer than I expected, and the early versions were embarrassing in ways I won't detail here. By the time it was working the way I needed, I had built something I realized I couldn't be the only one who needed.
The realization
Everyone doing serious writing at volume runs into some version of this problem. Agency owners producing content across dozens of clients. Content leaders managing writers, workflows, and deadlines. Founders who write to build their audience and reputation and don't have six hours a day to do it manually. Political writers, ghostwriters, reputation professionals.
They all need voice that actually holds. Accuracy that isn't embarrassing. A process that doesn't require them to become prompt engineers.
What I had built for myself should exist for them. That's where Ghosts came from.
What Ghosts is
I'll be direct about the stake here: I built Ghosts, I run Ghosts, and I think it's the right tool for this work. That's my view, not a neutral market survey.
Ghosts does what my internal tooling did, designed for people who are not me and didn't spend a year building AI pipelines by hand. It handles SEO content, online reputation management, political writing, ghostwriting for executives and high-profile people, marketing copy, and social media. The same multi-agent structure runs underneath: planning, research, fact-checking, and editing as separate steps handled by specialized agents.
It is not a technical system. You don't watch training videos to operate it. You don't write prompts. You bring your idea and your angle; the system pulls the rest from there. It writes in your voice, or whichever voice each job requires, because voice is built into how each assignment is approached, not applied afterward from a settings menu.
The AI is doing real work underneath. I'm not hiding that. The point is that the human stays in charge of the thinking while the system handles the production.
I spent a year doing this work at a volume that forced me to solve the problem for real. The solution I needed didn't exist, so I built it. Now it does.
If you do serious writing at scale and you've been let down by tools too generic for what your work actually requires, that's who Ghosts is for.
See how Ghosts works, or start writing with it.