Brian Alvey, the CTO of WordPress VIP, recently described what it takes to hand your work off to an AI: you have to document everything you know. Every preference, every standard, every decision you'd make if you were in the room. He called these documents "obsidian vaults for your agent harness" — the context layer that lets a bot operate your business instead of just answering questions about it.
He's theorizing. We built it.
What's in the Vault
We built The Charlotte Mercury as a system — and that system runs on a set of plain-text files that define how the entire operation works. There is a file that tells an AI collaborator who Jack Beckett is — not just his name and beat, but the words he uses, the words he never uses, how he opens a piece, how he closes one, what he sounds like when he's covering a zoning meeting versus a police log. There are files for every author, every beat, every publication. There is a master document that holds the institutional knowledge of the company — terms, people, preferences, decisions we've made and why we made them.
None of this is code. It's markdown. Plain text with structure.
The Documents Are the Company
Most people building with AI start with a prompt. A paragraph or two that tells the model what to do. That works for a single task. It does not work for an operation that publishes across three newspapers with distinct voices, distinct beats, and distinct audiences.
The difference between a prompt and a vault is the difference between giving someone directions to your house and handing them the keys. A prompt gets you one trip. A vault gives an AI collaborator enough context to do the work the way you would do it — to know that Beckett doesn't use passive voice, that The Farmington Mercury covers a town of fewer than 30,000 people in Connecticut, that a fact-check failure on a city council member's name is not a minor error.
When the vault is right, an AI session picks up where the last one left off. When it's wrong — when a detail is outdated or a voice rule is missing — the output drifts, and we catch it, and we fix the file. The document improves. The next session is better.
The Moat Nobody Talks About
The conversation about AI in media is fixated on which model to use. GPT or Claude or Gemini or whatever ships next quarter. That's the wrong conversation. Models are commodities. They improve on a schedule none of us control, and they're available to everyone at the same price.
The thing that isn't a commodity is what you feed them.
Two publishers can use the same model, the same tools, the same infrastructure. One of them has a vault — years of editorial decisions, voice rules, beat research, production records — encoded into documents the AI can actually follow. The other has a prompt. The output will not be the same. It won't be close.
Alvey sees this clearly. He talks about documenting "the soul of the company" so that every micro-task can be handled by whoever or whatever is best suited for it. The soul of The Charlotte Mercury lives in a folder of markdown files that a year ago didn't exist. Every session makes them more precise, more complete, more valuable.
That folder is not documentation about the company. It is the company — the part that makes the work ours instead of generic.
The model is the engine. The vault is the driver.