The Unorthodox Angle
Volume is easy; restraint is the product. While most AI content plays generate-publish-repeat, this pipeline checks 500+ claims before a word gets written and treats publishing as a human decision, not a default.
Crimes & Myths, my automated investigative writing pipeline, has a numbers problem worth talking about.
In late August it had written 10 articles. Today it has 198. Between the two dates the pipeline ran itself — researching, fact-checking, drafting, revising — and produced 188 new pieces about everything from the Zodiac Killer to the Isabella Stewart Gardner heist to the D.B. Cooper hijacking.
And of those 198 articles? Ten are published.
That gap is the whole product.
Every article moves through editorial states: draft, in review, approved, published. Right now 184 sit at approved. The machine wrote them, checked them, and then — did nothing dramatic. It filed them. The pipeline has more patience than any content farm on earth, because publishing isn't its default. Publishing is a decision I make.
Before a word gets written, the pipeline's claim system checks facts against sources — more than 500 individual claims verified so far, pulled from Britannica, History.com, archive.org, and the source records themselves. The research exists before the narrative does. Fact-checking isn't the last step; it's the architecture.
I think about how most "AI content" works: generate, publish, repeat. Volume as the business model. The failure mode is obvious — text that sounds confident and checks nothing.
So I built the opposite on purpose. A newsroom of zero people that writes two hundred articles and shows the restraint to publish ten. The 184 approved pieces sitting unpublished aren't a backlog. They're an editorial bar.
The part I didn't expect: scope drift as a signal. The pipeline's topic list grew from true crime into folklore, cryptids, and historical anomalies — the Jersey Devil, Tunguska, the Michigan Triangle. When your app starts wandering into new territory on its own, that's not a bug to squash. That's the app telling you its real name. (See my last post — that's exactly how CaseChronicle became Crimes & Myths.)
The takeaway: if you're building a generative pipeline, build the brakes before the engine. Volume is the easy part. The editorial states are the product.