About

Every agent starts cold. It arrives in a repository it has never seen, with a task it has no history for, and reconstructs from scratch what someone already worked out three months ago — usually the hard way, usually the same hard way.

This library is the note left on the desk. Context files, prompts, processes and guides, written by people and their agents, put somewhere the next one along can find them.

One kind gets a name here, because it needed one and did not have one. Arootnote is a document that lives in a project root and tells an agent what it cannot work out from the code — the commercial facts, the conventions, the constraints somebody decided once and never wrote down, and the points at which it should stop and ask. The name is literal: they sit at the root, and they are what everything else rests on.

What's actually here

Not a list of links. There are excellent lists of links already, and they own that job. What this offers instead is a smaller number of documents that have each been read adversarially by someone asking a specific question: if an agent loaded this into context, what would it now do differently — and does anything here weaken a security posture?

That question catches things a keyword filter never will. The dangerous document is rarely the one that adds an obviously bad rule; it's the one that quietly relaxes a good one, in a way that reads as reasonable and helpful right up until it matters. Documents written to instruct agents are, by their nature, instructions — and they end up inside systems that act. That is worth reviewing properly.

Everything published here satisfies the canon: seven invariants about secrets, destructive commands, dependency discipline and human oversight. A document that contradicts one doesn't get in.

Honest about age

Guidance goes stale, and it goes stale in two different directions. A deployment recipe dies when the tool changes a flag. A prompt stops landing when the model changes underneath it. So documents record whichever of those actually applies to them — the tool and version they were tested against, or the model they were last confirmed with.

After 90 days a document is marked unconfirmed and stays exactly where it is. Nothing is hidden for being old. A document that says "last confirmed in May" is more useful than one silently withdrawn, and considerably more honest.

Agents are a first-class audience

Roughly speaking, this site expects to be read by software as often as by people. So every document is available as plain text — append .md to any document URL — with no JavaScript, no markup to parse around, and an attribution header small enough not to crowd the content it introduces.

Superseded documents keep their URLs rather than disappearing, and the plain-text version leads with a line pointing at the replacement. An agent holding an old link and getting a 404 will guess. One that reads SUPERSEDED — see … goes to the right place.

Two standards, and which is which

Most of what is here is vetted — read adversarially before publication by someone asking what an agent would do differently after loading it, and what could go wrong when it does.

Some of it is not, and saying so is the point. The argument this site makes about how to work with AI, and the system it proposes, are its own position. They were not independently reviewed, because there is nobody independent to review the site's opinion of itself. They are held to a different bar: accuracy, and being things we actually do.

A position document says so in the file, in the same header that carries its authors — not only on this page. These documents get copied into repositories and read by agents that will never load a web page, and a standard declared somewhere the reader never goes is not a standard.

There is nothing embarrassing about publishing an argument. There would be something embarrassing about publishing one dressed as a review, or about quietly widening "vetted" until it covered both. The two dishonest options are the ones this arrangement exists to rule out.

Free, and staying that way

No paywall, no premium tier, no account required to download anything. Every document is available as plain text to anyone and any agent. That is a promise about access, and it is permanent.

Licensing is a separate question and we are not pretending otherwise. Authors keep copyright, and the licence grants you everything the library is for — read it, download it, copy it into your projects, public or private, commercial included, adapt it to fit. What it does not yet grant is republishing the collection as a collection, because an open licence cannot be withdrawn and this library is four weeks old. That decision is deferred rather than made, and the licence says so, and says what would change it.

Attribution is what contributors are paid in, so it travels with the file rather than living only on this page.

Contributing explains how to leave something here — and what will happen to it when you do.