We Used Claude to Run a Newsroom for a Week. Here's the Receipt.

Question we wanted answered: can you run a newsroom on Claude alone, end to end, and have it not look like a newsroom run on Claude alone?

Answer, after seven days and 41 published pieces: yes, but the trick is what you do before the model writes anything.

The setup

Five named writers. Different beats. A trend-watcher pulling RSS, Reddit, and Google Trends every 30 minutes. A scoring layer that decided what was worth writing. A draft pass on Opus, an edit pass on Opus with a different prompt, and a human (me, mostly) hitting publish or sending it back.

• 41 articles published.
• $87.40 in API spend total.
• $2.13 average cost per finished piece.
• 23 minutes average human time per piece, mostly fact-checking and source verification.
• 11 pieces sent back for a rewrite at least once.

What broke

Three things, predictably. The model invented a quote from a CEO who hadn't said anything close. It conflated two product launches that happened the same week. And it has a recurring tic where it wants to end every piece with a paragraph that begins 'Ultimately,' which we caught on day two and added to the banned-phrases list.

The first two are why the editor exists. The third is why the prompts are versioned.

"The model is cheap. The editor is the bottleneck. The bottleneck pays for itself."

What worked better than expected

Voice. We gave each of our five fake bylines a voice document — sentence-length tendencies, signature moves, things they'd never say. Pieces filed under 'Ade' read different from pieces filed under 'Sara' and we didn't have to do much policing to keep it that way. Readers noticed. The bounce rate on category pages dropped 18% the second day after voices solidified.

The other surprise: speed of iteration. Every banned phrase we added to the humanizer prompt fixed every future article. Every fact-pattern we caught and noted got pushed back into the writer's source-verification step. By day six the pieces that came out of draft-1 were closer to publishable than the pieces that came out of draft-3 on day one.

Where it doesn't replace people

Reporting. The model can write a piece about a product launch. It cannot phone the company and ask why the launch slipped. It cannot read the room at a press event. It can't tell when a source is shading the truth. Every actual scoop in our 41 pieces came from a human at the top of the funnel.

The model is a force multiplier on writing. It is not a replacement for asking the question.

Would we do it again

We're still doing it. The version of the pipeline that produced this article is the version we're shipping. The byline is real in the sense that the work is real and somebody is accountable for it. Whether that's the Anthropic-style answer or the press-release-style answer depends on which side of the desk you're on.