Today I stumbled across a video by Marie HaynesAn SEO consultant and AI-search analyst who has spent years decoding how Google ranks and surfaces content. She runs Marie Haynes Consulting and publishes prolifically on search and how AI is reshaping discovery.mariehaynes.comYouTube, @Marie_Haynes, Google’s OKF - The New Way to Structure Your Knowledge for Agents. I was intrigued by the premise.
Open Knowledge Format (OKF) is Google’s new, vendor-neutral way to represent knowledge such as the metadata, context, and curated insight that surrounds your data, so that AI agents can read it. The spec is actually quite simple. Mostly, it’s just a folder of markdown files with frontmatter conforming to some rules, an index.md, and a log.md. They call this a “bundle”. Their repo also contains some rules to help create, visualise, and search the contents of the bundles.
The idea traces back to 
Andrej KarpathyA founding member of OpenAI and former director of AI at Tesla, known for his research on neural networks and his widely-followed teaching on deep learning. He is the one who floated the “LLM wiki” idea.Wikipedia, Andrej KarpathyThe LLM wiki gist, who floated the notion of an “LLM wiki”. It’s basically OKF just using different words. Google’s version mostly pins down a few rules for frontmatter and filenames.
OKF seemed like a reasonable idea: link knowledge together in files, add a few rules, and make the whole thing easier for agents to read.
But I wasn’t far into the video before it hit me:
That’s just HTML. And HTML does it about 1000× better!
The Semantic Web called. It wants its idea back.
We had a name for this idea already: the semantic web.
Back in 2001, 
Tim Berners-LeeInventor of the World Wide Web, which he proposed in 1989 at CERN. He directs the W3C and set out the original vision of the Semantic Web.Wikipedia, Tim Berners-LeeSemantic Web road map (1998)Web Architecture from 50,000 feet (1998) pitched a web of linked, machine-readable meaning for software agents to traverse, with structured data embedded right in the documents. He had sketched it years before that, from “a level of 20,000ft”, in his Semantic Web road map and a companion paper, Web Architecture from 50,000 feet. Both make the same point: the Web was always meant for machines to take part, not just humans to read. OKF is that same dream, just with a new acronym.
That 2001 vision leaned on formal, machine-readable languages: RDF, OWL, SPARQL. They made sense when software needed explicitly encoded meaning. LLMs change that tradeoff. A model can read raw HTML, or raw anything, and infer a lot of the structure itself. The formalisms still matter for deterministic scripts and parsers, but for an agent reading knowledge, the page is often enough.
Every capability a knowledge format needs is already in HTML.
- Links - local or across the web. The link graph is the knowledge graph.
- Structure and style - that’s HTML and CSS.
- Machine-readable metadata -
<meta>,<link rel>, schema.org, JSON-LD, embedded in the page and read by crawlers for a decade. - A queryable data model - the DOM. One API, no SDK.
- Its own languages - JavaScript and CSS.
- Forms - state-changing actions, declared right in the markup.
- Auth and security - sessions, OAuth, CSP, CORS, same-origin.
- Accessibility - screen readers and keyboard navigation.
- Readable and writable by agents and humans alike.
- Local or hosted - opens from a file or a server, caches offline.
- Universal - every device ships a renderer.
- Backwards-compatible forever - HTML from 1995 still renders.
- Indexable - decades of search tooling.
I may also be biased. After all, I’m building an HTML-first web framework, JST.
Every OKF concept already has an HTML equivalent
The OKF concepts map almost directly onto existing web concepts:
| OKF / LLM-wiki concept | Native HTML equivalent |
|---|---|
A directory of .md files |
A directory of .html files, open in a browser, zip it, or serve it |
YAML frontmatter (type, tags, timestamp) |
<meta> tags in <head> |
| Markdown body | <body> |
Markdown links [x](/path.md) |
<a href="/path.html"> |
index.md (progressive disclosure) |
sitemap.html, the rarely-seen-but-once-ubiquitous sitemap |
log.md (update history) |
log.html |
“Concept ID” (path minus .md) |
The document’s own URI, like file://…/path or https://host/path |
# Citations |
<cite> / <a> |
Their bespoke viz.html graph viewer |
…the browser itself |
OKF, but HTML-ified
Mostly as a joke, but also to prove my point, Claude, Codex, and I rewrote the entire Google OKF repo and replaced every Markdown concept with its HTML equivalent: the spec, the README, the example bundles, the lot. You can read the result here. The goal was to leave the documents as untouched as possible.
The diff converts all 78 example documents to HTML, and every cross-link still resolves.
We deleted around 59,000 lines of bespoke tooling: the Python “enrichment agent”, the TypeScript toolbox, and the custom graph viewer. HTML brings its own viewer: the browser. Open any sitemap.html from disk and you’re browsing the knowledge base. No server, no install, no build step.
So…
OKF accidentally proves HTML was the right abstraction all along.
Just use HTML!
If HTML isn’t human-writable, get your agent to write it. lol.