Between 2023 and 2026 I read some three thousand AI white papers. I highlighted them, excerpted them into Word, then moved the excerpts into Obsidian, where I could tag and cross-link them and use Claude to research and write from them. The vault became the most useful thing I owned, and it was private. Inquiring Lines is that vault rebuilt as a public site, so that anyone can explore the research by topic and, above all, by question.
The problem with categories
Research libraries are usually navigated by category, by citation, or by keyword. Each fails a reader in a particular way. Categories put a paper in one place, while most papers bear on several questions at once. Citations connect papers that know about each other, and miss the ones that don't. Keywords find the vocabulary of a field and miss the same idea expressed in another field's terms.
Researchers, and curious readers, don't think in categories. They think in questions: does a model behave differently when it knows it's being tested? Can AI judge what research matters? Questions like these cut across fields. The best evidence for one might sit in a paper on interpretability, another on reinforcement learning, and a third on the psychology of being observed. A taxonomy separates them; a shared research interest connects them.
Content generated, not just stored
The site is built from layers of content, most of it generated with AI and checked along the way:
- Excerpts. More than 2,400 papers and web sources, each excerpted verbatim. A model chooses which passages to keep (abstract, introduction, method, discussion, conclusion); it never paraphrases them.
- Synthesis notes. About 2,400 notes, each stating one claim in a sentence-length title, citing the papers behind it, and linking to the notes it extends or contradicts.
- Topics and clusters. Some 90 research topics, and clusters and themes computed from the notes themselves.
- Inquiring lines. About 10,000 of them, four times as many as there are papers. Each line is a research question written in plain language, with an interpretation of how the library answers it and the notes and papers it draws on.
The lines are the heart of the site. They outnumber the papers because the same paper bears on many questions, and the same question is answered from many papers. Generating them, classifying them, clustering them into areas and writing plain-language versions took more work than everything else combined, and it is what makes the library navigable.
Navigation that maps to the content
Because the lines are questions, the navigation can be too. A reader can:
- Follow a question to research from different fields that bears on it, rather than to the papers filed under one topic.
- Search by meaning across the whole library: papers, notes, topics and lines, ranked by what a sentence means rather than which words it shares.
- Start from a paper or a post. Paste an arXiv link, or a social post about AI, into Match, and see the lines of inquiry and research it bears on.
- Browse up and down from broad areas to themes to individual lines, each level labeled in plain words.
- Follow a story. Inquiries take an AI news story, such as a frontier model escaping its sandbox, or the debate over whether alignment still works, and follow it into the research behind it, with a timeline, curated research directions, and the lines a curator picked under each.
The system behind it
I designed and built the whole system with Claude. New research moves through a pipeline: papers are taken in from arXiv or the web, excerpted, carved into sources, and given library context; a model writes the synthesis notes; validation checks every note's format, citations and links; a connection pass links each note into the library; a reader layer writes plain-language versions; and new lines of inquiry are generated, classified and clustered. A render step publishes some 14,000 static pages, which go to a staging site before they reach the public one.
Around the content sits an admin console of my own design: a front-page stream for featuring papers, briefs and one-off features, a queue of incoming papers, a page of drafted social posts for every line, an activity console that separates people from crawlers, and an Inquiries tab where each inquiry is held until I choose to publish it.
Editorial guardrails matter as much as the machinery. Excerpts stay verbatim. Every source behind an Inquiry is opened and checked before it is cited. Pages say where sources disagree, and each Inquiry ends with a note on how it was made, by AI, under a human curator.
What it shows
Inquiring Lines is a case study in three things I now do for clients. It uses AI to generate original content at scale, with the checks that make it trustworthy. It designs navigation around what the content actually is, in this case questions, rather than forcing it into a standard taxonomy. And it treats publishing as a system: pipeline, staging, admin tools and analytics, so a small team, or one person, can keep a large site alive. If you have a body of research, a catalog, a course or an archive that people struggle to find their way through, this is the kind of system I can design and build. Ways to work together →