Field Note: Open Knowledge Format
This week Google Cloud released the Open Knowledge Format. It’s plain markdown files with a short labelled header — a few fields like type, tags and a link (they call it YAML frontmatter) sitting above the text. It isn’t tied to any model, framework or company: anyone can write it and anything can read it, from a web server or an LLM to Obsidian, Notion, or cat in a terminal. It’s also, more or less, how I’ve kept my own notes for years.
What’s nice is that it doesn’t put a company between you and your own knowledge. You can read it without an SDK, keep it in git, see what changed, and copy the whole lot somewhere else whenever you like. A metadata store you reach through an API doesn’t give you that: if the terms change, you can lose access to your own material.
I’d still be careful. “Open” from a big cloud vendor is always worth a second look. What reassures me is that there isn’t much to lock in here, because they really are just files; if your notes are already in markdown, you’ve mostly got it already. The thing to watch is the next year or two, and whether the format stays simple and the tooling stays optional rather than the useful parts slowly ending up only on Google’s cloud.
Where is your own knowledge kept right now?
Also this week
Microsoft’s “AutoJack” research. Microsoft showed a proof-of-concept (since patched, and never in a shipped release) where a malicious web page reached an AI agent’s unprotected local service and ran a command on the machine. The lesson holds even though the bug is fixed: once an agent can browse the web and also reach services on your own machine, “it’s only local” stops being a safety guarantee. Worth a read if you run local tools for your agents.
Turning a prompt into a reusable skill. I used to paste the same prompt in by hand until I turned it into a skill: a small file the model loads itself when it’s relevant. Anthropic’s skill-creator does that for you — describe the prompt, or hand it a whole conversation, and it writes the skill. The catch is not making too many, or you end up with a pile you can’t keep track of.
Cloudflare’s temporary accounts. I lean on Cloudflare for a fair amount of my own work — a couple of my sites run on it (this newsletter’s included), along with a few small tools, and its headless browser does a lot of my page-reading when I’m researching. So this one caught my eye. wrangler deploy --temporary now puts a Worker live with no sign-up and nobody logged in; a human then has 60 minutes to claim it, or it deletes itself. Useful if you run things unattended, and a fair question about which steps you still want a person to approve.
AI and ungoverned knowledge. Gartner expects companies to drop 60% of their AI projects by the end of 2026 for lack of “AI-ready data,” and Deloitte’s 2026 figures put governance well behind ambition. It’s what note-takers already knew, now arriving in the enterprise: the hard part was rarely the model; it was usually the quality of what you give it.
— Jim