My AI runs on a folder of markdown. Google just standardized it.

Most of what runs my work is a folder of text files. Plain markdown, nothing fancy. Notes on each project, a running status file that says where things stand, an audit file a build agent reads before it's allowed to ship anything. The tools read those files before they act and write back to them after. Somewhere in the last year I started talking to my own software through documents more than through chat windows, and I never decided to do it. It piled up one file at a time, because it worked.

I figured this was just a me thing, the kind of habit you don't bring up because it sounds like hoarding. Then last week Google published a spec that describes the exact setup, gave it a name, and stuck a version number on it.

So it has a name

It's the Open Knowledge Format, OKF for short, and the whole thing fits in a sentence. A folder of markdown files, each one describing a single thing (a table, a metric, a process, whatever you need to write down), a few structured fields at the top, and plain links between the files so the folder reads like a little wiki. That's it. No database, no API key, no account. If you can write a text file, you can write OKF.

What matters is what the agent does with it. Instead of searching the same scattered docs every time it needs context, it reads a folder that grows as you go, and it can update the files itself. The knowledge builds up instead of getting reassembled from nothing on every request. Andrej Karpathy described this pattern a while back and put it better than I will: the bookkeeping that makes people give up on their personal wikis, all the cross-referencing and tidying, is exactly what a model is good at. It doesn't get bored. So you let it keep the notes.

Google Cloud shipped OKF on June 12. Version 0.1.

What it isn't

Here's where I think most people are about to go wrong. Google, plus AI, plus open, and every marketer reaches for the SEO playbook. They'll assume OKF is a new file you bolt onto your website so the AI engines rank you or quote you. It's not. OKF came out of Google's data team. It's about feeding your own knowledge to your own agents inside your own systems. It has nothing to do with whether ChatGPT or Google's AI answers ever mention your business.

We've seen this one before. llms.txt did the rounds not long ago, a little file you'd drop on your site to feed the language models. Everyone added it. Then Google said flatly that it does nothing for search. OKF sits even further from an SEO trick than that did. If someone tells you adding OKF will get you into AI Overviews, they haven't read it. Don't put it on your marketing site expecting something to happen, because nothing will.

Where it matters

It matters at the knowledge layer, the part of AI work nobody puts on a slide. Everyone rents the same models. You and your competitor can buy the same frontier model this afternoon, so the model was never going to be your edge. The edge is whether your knowledge is in a shape the machine can use.

Mostly it isn't. It's smeared across a CRM, a wiki nobody updates, a shared drive, a few Slack threads, and the head of the one person who knows how the billing logic works. Point an agent at that and it rebuilds the same answer from scratch every time, wrong in a new way on each pass. OKF is one fix for the mess: put the knowledge in plain files the agent can read, and the next agent, and the one after that, with no custom integration for each. I've run a homemade version of this for a while. The audits I keep mentioning are part of it, a set of files a build agent checks before it touches anything. Mine was held together with tape. A real standard underneath it is an upgrade I'll take.

Now the caveat. OKF is two days old. It's version 0.1, and Google calls it a starting point, not a finished standard. Every example in the announcement is a database table, which tells you which room it was born in. Don't rebuild your operation around it this week. Hold onto the pattern, not the version number: plain files, readable by people and machines, that your agents read and keep current. That predates OKF and will outlive whatever OKF becomes.

Legible wins

This was always going to get standardized, because it works and because it leans on exactly what the machines are good at. The companies that pull ahead in the next stretch won't be the ones with a better model than you. They'll be the ones whose knowledge the model can read.

My folder of markdown was never clever. It's legible, to me and to the tools I hand the work to, and that turns out to be most of the game. I've spent a year writing documentation for an audience of robots and feeling a little weird about it. Now it's got a spec number. I'll take the validation.