Recipes: research, decisions, and reading

A folder named "to read" holds forty paper PDFs, a few of them scans. This lesson wires the parts you already know into three recipes for folders like that: a research workspace, a meeting-and-decision log, and a reading log. What changes between them is only which panels you lean on.

All three run on the same machinery. Markdown goes in, whether you wrote it or a live feed brought it, and background extraction builds the knowledge graph. Curation panels keep that graph accurate, and every answer arrives with the source sentence behind it.

Recipe 1: a research workspace from a folder of papers

Back to that folder, plus the dozen article links you keep meaning to read. The goal: ask questions that reach across all forty, like "who else built on this method?", instead of reopening PDFs one at a time. Your originals never change, because the app reads a copy and organizes what's inside.

You don't have to wait in silence. Send a chat message mid-build and the "Knowledge graph still building" dialog pops up, since an answer right now would search a half-finished graph.

It offers "Send when extraction finishes" or "Send now", and whatever you pick is remembered for the rest of the build. Held messages wait in a strip above the composer, each with "Send without waiting", "Edit and requeue", and "Remove". Saving a single note never gates like this; it only shows a quiet hint.

Expected outcome. Now you can ask questions that span papers, because the graph lets the AI follow connections from note to note until it reaches a complete answer. On any entity page, hover a connection chip and the exact sentence behind the fact appears, cited as "{note} · line {n}".

• "Topic map" in the left rail: named themes, each with a summary.
• "Summary note": pulls a theme's notes and concepts into one hub note (a map of content, or MOC).
• "Map mode" in the view settings: your themes spread out like continents on a map.

The refinement loop. That auto-built vocabulary is only a draft. Open "Domain vocabulary" in the left rail, press "Teach the AI your field", tweak the predicates it learned, and Save. To bring the forty papers you already imported up to the new vocabulary, run the Extraction panel's "Re-extract everything", or chain the whole loop with the Lifecycle panel's one-click "Tune up AI search": Domain vocabulary → Re-extraction → Synonym check → Relationship check → Topic map.

Recipe 2: a meeting log that remembers what you decided

The setup: weekly product meetings, one markdown note each. Six months later, someone asks: "what did we decide about the Atlas launch, and when did it change?"

Folders and tags can't answer that. They file your notes; they don't read them. The statement layer can, because it caught your decisions verbatim as you wrote them.

With "Capture your decisions and questions" on (it's on by default), extraction saves the author's own sentences, word for word, as statements: decisions, open questions, claims, quotes, plus standing preferences, values, stances. Each one is tied to the entities it's about, a quote from someone else carries its attribution, and a stance wears a favor/against/neutral polarity badge. On every entity page they show up under "In your words", grouped by kind.

meetings/product sync (March).md
  "Meeting notes, March 4, 2026."
  "We'll launch Atlas in early June."

meetings/product sync (May).md
  "Meeting notes, May 12, 2026."
  "Decision: the Atlas launch moves to September."

Revisions panel, later:
  NEW      "Decision: the Atlas launch moves to September."
  REVISES  "We'll launch Atlas in early June."
  product sync (May) → product sync (March)

Expected outcome. On the Atlas page, the March decision shows struck through, with a "Superseded" badge and a "Superseded by {note}" link that opens the May note. You can answer "what did we decide, and when did it change?" straight from a readable trail built out of your own sentences.

Recipe 3: a reading log that compounds

The setup: books as epubs, long-form articles as URLs, the occasional lecture recording. You want your reading to build into something connected: which authors keep reappearing, which ideas bridge two books.

Expected outcome. A reading workspace that answers back: the entity table shows who and what dominates your reading, and "Show on canvas" throws exactly the filtered rows you're looking at onto the 3D graph. Nothing here is a snapshot. Collections re-evaluate against the live graph every time they render, so the counts move as your notes do.

One pipeline, three stories

Strip away the story and the three recipes are the same five moves. Once you see the pattern, you can improvise your own recipe for any domain: a job search, a home renovation, a thesis.

Wrap-up

• Research workspace: wizard in, five stages run, answers that cite the source sentence out. Then refine with "Domain vocabulary" and re-extraction.
• Decision log: dated notes and clear sentences in; "In your words" and "Revisions" give you a self-maintaining trail of what changed and when.
• Reading log: books and articles in; the entity table and living collections turn reading into a network that compounds.
• Same machinery every time: capture → teach → curate → explore → ask.

Next, the same five moves run a full investigation, from a pile of scattered sources to a conclusion you can cite.