Skip to content
All logs

Structure first: what 8,800 cybersecurity topics taught me

Cybersecurity knowledge is scattered across blog posts, slide decks and tool READMEs. Two years of restructuring Cyber Library taught me one thing: the hard part isn't writing, it's the shape.

Florian Amette5 min read

Learning cybersecurity feels like reading a library where every book has its own index.

A blog post here, a slide deck there, a tool README, a forum thread from 2014. Each one is useful. None of them agree on where things go. The same technique gets three names, and the thing you actually need sits between two resources that never link to each other.

That's what frustrates me. The content exists. The shape doesn't.

So in November 2024 I started Cyber Library: one tree for the whole field, in open source. Two years later, the writing turned out to be the easy part.

Version one didn't scale

The first README promised a "structured, up-to-date repository of cybersecurity knowledge" in eight categories. The files looked like this:

offensive/network/enumeration/memcache_enumeration/service_detection/identifying_open_memcache_ports_default_11211.mdx

A path that tries to say everything says nothing. And every new page raised the same question: where does this go?

Here's a short history of that question, straight from git:

WhenWhat happened
Nov 2024Six top folders, including concepts and technology
Feb 2025concepts becomes foundation, technology disappears, intelligence appears
Jul 2025Web renamed to Browser. Then back to Web. Same day.
Dec 2025One taxonomy file split in three. The cyber one alone ends up near 1 MB of JSON
Mar 2026A full refactor: 522 files, 52,668 lines deleted. Seven pages survive
Oct 2026Cloud moves out of Server: 304 files, plus 301 redirects

Every row is a decision that seemed right when it was made.

The problem is the shape, not the content

Security knowledge doesn't have one natural axis. You can sort Active Directory attacks by kill-chain step, by protocol, by tool or by attack surface. Every option is defensible. Mixing them is what kills you: the day one folder sorts by step and its sibling sorts by protocol, every new page has two homes, and the tree starts to lie.

Scaling a knowledge base isn't about adding pages. It's about adding pages without moving the old ones.

The rules that finally held

These came from breaking things. Each one is now a written rule.

1. Structure first. Every page is planned as a topic in the taxonomy before anyone writes it. Today the tree holds 8,863 topics, and most of them are still drafts. That's the point: the map exists before the territory.

2. One sorting axis per level. Never mix criteria at the same depth. Active Directory is now sorted by attack surface, not by kill-chain step, and the catch-all "Enumeration" section is gone.

3. One concept, one place. If a topic could live in two branches, the tree is wrong, not the topic.

4. No junk drawers. Never a category called "Vulnerabilities". A category needs at least two children. "Windows" and "Linux" sit under "Operating Systems", not side by side at the top.

5. Name the leaf, not the path. It's Injection › Command, not Injection › Command Injection. The page about SMTP smuggling is titled Smuggling. The tree already says SMTP.

6. Stable URLs by design. A topic is a structure (a folder with no page), a section (a folder with its own page) or a page. A section keeps its URL when children arrive later, so growth doesn't break links.

7. "Ready" means it won't move. A branch is only marked ready once its subtree is stable enough to scale without reorganizing.

8. The structure is the source of truth. Not the folders. The taxonomy lives in a database, behind an MCP server with tools like topic_find_similar and topic_check_duplicate_title. The Markdown files follow it.

Where it stands

Topics planned8,863
Depth17 levels
Pages written1,793
Languages70
TracksFoundation, Offensive, Defensive, Governance, Intelligence, Career

And because a structured library is easy for machines to read too, it ships a free, read-only MCP server. Plug it into Claude or ChatGPT and your assistant can search the tree, read a page or walk an outline.

It's still not solved

I won't pretend the tree is finished. A few folders still have case-variant duplicates. Some rules contradict older ones. Every big move still means hundreds of files and a pile of redirects.

But the question "where does this go?" now has an answer before the page exists. That's the difference between a pile of notes and a library.

Read it, break it

If you find a technique in the wrong branch, open an issue. That's exactly the bug I want to hear about.

Ship the fix. Then share the patch.