Extracting and Scaling Intellectual Assets with AI

Ghosttown is an open-source project for extracting expert voices and reasoning patterns from public corpora — mailing lists, newsletters, LinkedIn posts. You open it in Claude Code, point it at an expert's writing, and get a Ghost you can talk to.

I just added John Cutler to the list of ghosts. Of course, it is nothing like John and his wisdom, just an echo of his tone. But what if we could go further and learn not only what experts say, but also how they reason about things? Would AI then be more helpful in making strategic decisions?

The easy part — a demo

Copying how someone talks is the easy part. It makes a nice demo. Here is what that looks like in Ghosttown. You activate a ghost and ask it anything:

/ghost-me trump

explain me this repo https://github.com/openclaw/openclaw

Trump:

Look, I know repos. I know GitHub, okay? Some people don't know GitHub — I know GitHub very well. The best people tell me about it.

So this OpenClaw — and it's a lobster, they use a lobster, which is actually very cute, I like it — it's a personal AI assistant. Runs on your own computer. Very smart, actually.

And you know what's beautiful about it? It's YOUR assistant. Not Google's assistant. Not some RIGGED system where they read all your messages. Yours. On your device. That's America First thinking right there, okay?

Funny. And it works. But it doesn't make a business. The real value is somewhere else.

The real prize

An expert's judgment lives in one person's head. It can't go beyond what their calendar allows, and it's gone the day they quit. You can't copy their time. But you can copy the way they reason — and that part can be shared.

This connects to something I've written about before: AI doesn't replace people, it replaces tasks — but only when the organization is designed to let expertise flow. A Ghost is one way to make that happen.

The Ghost

Picture the one engineer who understands the whole system. At 2am, or when ten people need them at once, they're not around. The Ghost is. It doesn't replace them. It covers the moments they can't be there.

The real question changes. Not "can it sound like me?" but "can it reason like me?" Show it an org chart or a tough decision. Does it see the problem the way the expert would? Does it spot what's broken, find where things get stuck, and say what to fix first?

The Intellectual Asset

Those thinking patterns are the real prize. We call them the Intellectual Asset — IA for short.

An IA isn't a fact about the company; it's a rule the expert carries from one job to the next. "We have 500 employees" is just a fact. "Once you pass about 300 people, decisions slow down unless you do X, Y, Z" is an IA — a rule that works on any company, not just this one. The expert's methods, rules of thumb, and mental models are the asset. The Ghost is just how you reach it.

This is also where routine and adaptive expertise diverge — what makes an expert irreplaceable isn't their ability to repeat a process, it's the judgment they've developed across contexts.

Extracting an IA takes four passes over the expert's writing: first, their rules of thumb (if X then Y); then the principles behind those rules; then their consistent biases and stances; finally, voice — how they frame and deliver it. Strip those four layers out and you have something that holds up on new problems, not just the ones it was trained on.

The test is simple: a Ghost built from the IA should tackle problems the expert was never asked about — and still reason the way they would. If it can't do that, it's a quote machine, not a Ghost.

These are my thoughts behind Ghosttown.

The ladder

There are four layers here, and most projects stop at one:

Layer	What it copies	Status
1 — Voice	How the person talks	✅ This repo
2 — Persona	How the person reasons	🟠 This repo (MVP)
3 — Intellectual Asset (IA)	Their methods, rules of thumb, mental models	Pipeline starts this
4 — Institutional Ghost	The IA plus a company's own data	Future

Ghosttown is the first step — a small, copyable example that shows the idea works. The four-step extraction process (rules of thumb → principles → biases → voice) is how we pull an expert's IA out of their writing.

The first sign it's working is how much it packs into an answer. If it says 2–4× more per word than plain Claude on the same question, it's really using the expert's judgment, not just copying the style. Craig Larman's ghost hits 3× in early evals — also available as a separate repository: Craig as a Service.

Toy with it, fork, add your content, expand: ghosttown on GitHub.

Be my ghost.