As of October 22, 2026: 3,006 papers and sources · 3,151 synthesis notes · 13,256 inquiring lines · 91 topics · 3 published Inquiries
The idea
Research libraries are usually sorted by category, citation or keyword. Each one fails a reader in its own way. A category puts a paper in one place, but most papers bear on several questions. Citations connect papers that know about each other and miss the ones that don't. Keywords find a field's vocabulary and miss the same idea in another field's words.
People who read research think in questions. Does a model act differently when it knows it's being tested? The best evidence can sit in a paper on interpretability, another on reinforcement learning, and a third on the psychology of being watched. A category keeps those papers apart. A question brings them together.
So Inquiring Lines is built around questions. You can still browse by topic, but the main way through the library is to follow a question.
Where it came from
Between 2023 and 2026 I read some three thousand AI papers. At first I highlighted them and copied excerpts into Word. Then I moved the excerpts into Obsidian, a notes app, where I filed them into about 90 topic files, each excerpt with the paper's title and link. It was faster to search than any AI doing a web search.
Inside the vault I used Claude to research and write. A plugin called arscontexta changed what the vault was. It reads across the topic files and writes a layer of notes above them: short syntheses of one idea each, linked to related ideas. When I wrote, Claude reached for those notes first, instead of searching every excerpt. The vault stopped being an archive and became a pipeline. New material comes in, gets turned into notes, gets connected to older notes, and older notes get revised when newer work sharpens or challenges them.
The vault also has an operating manual that Claude reads at the start of every session. When a way of working stopped working, I changed the manual. Notes that contradict each other are tracked as tensions, not smoothed over.
The vault was the most useful thing I owned, and only I could use it. In May 2026 I started turning it into a public site. It took the name Inquiring Lines and its own domain, inquiringlines.com, on June 6, 2026.
Why questions: where the idea comes from
Two parts of my own work led here. For years I built websites and content management systems for more than 20 companies. A content management system files things by type and category, because that's how the people who publish think. Readers rarely think that way. They arrive with something they want to know.
In the 2000s I worked on social interaction design: how people interact with each other through software. Much of that was about how people follow ideas and talk about them. People pick up a thread from someone else, add to it, argue with it, and pass it on. An idea travels as a conversation, not as a filed document.
Inquiring Lines puts the two together. It is a content system underneath. What readers move through is the shape of a conversation about research: questions, the evidence on each, where it disagrees, and the next question.
Three layers of content
The library is built in three layers. Each layer is a different kind of thing, and each one depends on the layer below it.
- Excerpts (3,006 papers and sources). Passages taken word for word from each paper: the abstract, introduction, method, discussion and conclusion. Of these, 2,291 are arXiv papers; the rest are lab posts, reports, journal articles and other web sources. AI chooses which passages to keep. It never rewrites them.
- Synthesis notes (3,151). Each note states one claim in a sentence-length title, cites the papers behind it, and links to the notes it extends or contradicts. Most are findings; others name patterns, tensions, open questions or methods.
- Inquiring lines (13,256). Each line is a research question in plain language, with a short framing, an account of how the library answers it, and the notes and papers it draws on.
There are more than four lines for every paper, because one paper bears on many questions. Generating, sorting and plainly rewording the lines was more work than everything else put together.
A line isn't a search result in the usual sense. It doesn't hand back one fact. It's an approach: a way of pursuing a question across the research. Here is one I like, as a designer: What workarounds do users build when AI interfaces don't match their mental models? It says honestly that the library has little direct research on workarounds, then shows what explains the mismatch.
Navigating by interest, not by keyword
Research is usually searched in its own vocabulary: model names, benchmarks, method terms. That works if you already speak it. Most people who care about AI don't, and even researchers come to a paper with an interest, not a keyword. So Inquiring Lines is navigated by simplified research questions, nested as inquiring lines, that put the research in terms of the interests people bring to it. Language models make that possible. They can read a paper and say what question it bears on in plain words, which lets navigation sit closer to people's interests than keyword search can.
This is related to, but not the same as, the interestingness I write about in the last chapter of Conversation Is the Medium. There, interestingness is the mutual context a person builds with an AI in conversation. Here, the AI's facility with language is put to work on navigation: browsing and searching research by what people are interested in. It shows up in four places.
- Lines are questions worth following, not summaries. A summary closes a topic. A line opens one and points to the next lines. Each line also carries one of four moves: it bridges fields, surfaces a tension, finds a pattern, or opens a frontier.
- Lines say when the research is thin. A line that admits a gap is more useful than one that pads it. It tells you where the open question is.
- Explore opens the threads in a post. It doesn't give one verdict on a post. It lists the questions the post raises and doesn't pursue. Each thread is rated from 1 to 5 for interest. A 5 means the library holds research the post doesn't mention, and that research changes how the thread reads. A 1 means obvious or thin. Threads rated 2 or lower are dropped, and the rest are ordered by interest.
- I tested for it. When I reviewed Explore's test pages, I marked each suggested question "new to me", "obvious" or "off". The marks asked whether a question was new to the reader, not only whether it was relevant.
Ways in: each feature, and how it works
Inquiring lines
What you see. The inquiring lines index is a map, not a list of 13,000 questions. Eight broad areas, such as Agent Safety and Oversight or Trust and Conversational Experience, open into 32 themes, each shown with a few sample questions. A line page shows the line's place in a chain of parent questions, so you can step up to a broader question or across to a related one.
How it works. Each synthesis note suggests a few questions it could help answer. Each question becomes a line: the library finds the notes closest in meaning, and Claude writes the answer from those notes and cites them. The themes come from the questions themselves, grouped by meaning, not from my original folders. That gives the site two taxonomies on purpose. Topics organize the sources. Themes organize the questions.
Synthesis notes
What you see. A note page states one claim and links to the papers it comes from, the notes it extends or contradicts, and the lines that use it as a source. Browse the notes.
How it works. Notes are written in the vault with the arscontexta method. Every note's format, citations and links are checked before it goes on the site. A plain-language layer then adds a short headline and the questions each note bears on.
Topics
What you see. 91 topics, such as personas and personality or frontier AI risk and recursive self-improvement. Each topic page holds its papers and notes.
How it works. Topics descend from the Obsidian topic files. A new paper is filed under one topic, which the system suggests and I confirm. My personal tags in Obsidian never reach the site.
Search
What you see. One search box that returns papers, notes, topics and lines, each in its own group. Try a question.
How it works. Semantic search: results are ranked by meaning, not shared words. Lines appear as their own kind of result, so a question can land you on an approach to it.
Explore: paste a post, get a map of research threads
What you see. On Explore, you paste a paragraph from a social or blog post, or a link to it. In up to a minute you get a page of the lines of inquiry the post opens. Each line shows where the research stands (supported, mixed, contested or early) and one to three papers or notes to start from. If the post names a paper that's in the library, the page says "Named in the post" and offers that paper first. More lines wait in a panel below, each named before you open it.
How it works. A language model reads the post and lists the questions it raises: its claims, the assumptions under them, ideas mentioned in passing, and the question it puts to its readers. Each question is rewritten four or five ways in the words researchers use, once from an opposing idea, so the search finds studies that complicate the post as well as ones that back it. It also names the research fields behind practitioner wording: a consultant's "owns the outcome" reaches research on delegation and supervisory control. The model then keeps items that fit the question, not just the words. A last check removes any sentence that names a number or a name its source doesn't contain.
Links work too. An arXiv link goes to Suggest a paper. An X post comes in through X's public embed. LinkedIn and Medium pages can't be read from outside, so the page asks you to paste the text, and I chose to keep it that way.
Suggest a paper
What you see. On Suggest a paper, you paste an arXiv link or ID. If the paper is already in the library, you go straight to it. If it isn't, you see the lines and research it connects to, and you can suggest it for the library. Each match gets a page you can share.
How it works. The match reads the paper's full text, not just the abstract, and compares it with the whole library by meaning. Since October 21 it gives newer papers a small boost when two matches are close. I kept the boost small on purpose. A suggestion goes into my intake queue; it isn't added automatically.
Inquiries: news stories followed into the research
What you see. An Inquiry takes an AI news story and follows it into the research. It has a timeline, the primary sources, where the sources disagree, curated directions into the library, and a note on how it was made. Each one is a dated snapshot, not a living page. Three are published under Featured: Frontier AI risk and recursive self-improvement, Correct but not understood and Who judges expert work when AI does the work?
How it works. First I check the library's coverage of the story, and new sources fill the gaps. Every source behind an Inquiry is opened and checked before it's cited. Then the system ranks candidate lines under each direction, and I tick the ones that belong. Inquiries are held by default. Nothing is public until I release it.
From paper to library entry
In product terms, a new source goes through six steps, with a check at each one.
- Choose. I start from the gap: the questions the library can't answer yet, often found by running Explore on a typical post. I read research organizations' own pages first, such as the labs, METR and Stanford HAI. This year's work comes first; older work only as a foundation current work cites.
- Excerpt. AI selects passages from the paper; the words stay the paper's own. Text from a broken PDF is flagged for me to review, never used as is.
- Write notes. Synthesis notes are written from the excerpts and must pass the format, citation and link checks.
- Connect. New notes are linked to related notes in the library. Older notes get revisited too. On October 22, a pass added 403 links across 342 older notes. A check kept 403 of the 1,435 links proposed.
- Make lines. New questions become lines and join the nearest existing theme. When enough new questions stop fitting, the themes are redrawn, so the map of questions keeps up with the research instead of forcing new topics into old boxes.
- Publish. Everything goes to a staging site first. It reaches the public site only when I say so, and the previous version is kept to roll back to.
Checks catch real mistakes. One line said "most rehired more than half"; the source's figure was 35.6%. A note on workers hiding their AI use said about 70% hid it; the source said 69% mentioned stigma. Both were corrected in the live text.
The editorial side
An admin side, which I designed, lets one person run the site.
- The front page is hand-featured. A list called "Suggested for the front page" ranks papers I haven't featured by recency, Hugging Face upvotes, a match against this week's AI headlines, how many lines and notes already lean on it, and a few other signals. The list is an aid. I choose what goes on the page.
- Posts. Every line has a drafted social post, with search, a date filter and a way to hide the ones I've used. Posts name their sources. Share links show which posts bring readers.
- Inquiries are held until I preview and release them.
- Custom content management. Papers enter an intake pipeline, and papers that visitors match or suggest are queued for intake.
- Engagement dashboards. An activity console separates people from crawlers and follows what readers do with each feature, in far more detail than standard analytics.
Attempts and changes over time
Most features took several tries. The dated record below comes from the site's history.
- Late May 2026: the first 5,588 lines. The first lines were generated from the questions in the notes. At first I sorted them by four kinds of move, such as "surfaces tensions". On June 1 I reorganized the index by areas and themes drawn from the questions themselves. The moves stayed as a filter.
- June 2: thin papers fixed. I added verbatim passages from the PDFs to 447 papers that had only an abstract. Search started returning lines as their own kind of result.
- July 14: the first Explore. It was a walk through areas of questions, not a tool for posts. That walk is still on the site as a browse view. Line pages started leading with a plain headline above the question.
- September 28: a new front page. The front page became an editor's stream of featured papers, each shown with its lines, notes and related papers.
- October 6: Inquiries. The first Inquiry, on frontier AI risk and recursive self-improvement, went live with 40 new sources behind it. From October 8, Inquiries were held until I released them.
- October 13 to 16: Explore for posts. Four rounds of testing on real LinkedIn and Substack posts, judged item by item. Round 2 raised relevant or partly relevant results from 60% to 77%. Round 3 wrote full answer pages and found one of the seven threads I had listed for a test essay; round 4 read the post for its threads first and found four or five. Then I changed the goal: Explore became an orientation map, not an answer. It went live for visitors on October 16.
- October 18 to 21: Explore grows. A summary of the post's argument, the research fields behind practitioner wording, the "more lines" panel, links as well as pasted text, and "Named in the post". That last one came from a real case: a map that cited a paper's notes but never the paper the post was about.
- October 19: front-page suggestions. The Suggested list and the news match came after the front page had shown the same ten seed papers for weeks, because nothing new had been featured.
- October 22: today's library. Three intake batches went live: research on agentic commerce; 76 sources of design-community research on interaction design and UX for AI, from CHI and CSCW papers to guidance from NN/g, Microsoft Design and IBM; and research filling gaps in an essay I'm writing. With the pass that linked older notes to newer ones, the library reached 3,006 papers.
What it shows
Inquiring Lines shows three things I can do for a team. It uses AI to generate original content at scale, with checks that make it trustworthy. It designs navigation around what the content is, here questions. And it treats publishing as a system one person can keep current.
It also shows how I work with AI as a designer: most decisions on this page came from looking at real output and changing course. If you have a body of research, a catalog, a course or an archive that people struggle to move through, I'd like to talk. Ways to work together →