How to Organize Research for a Book: A Step-by-Step Guide

What you walk away with

• Research organized by the questions the book asks, not by the format the material arrived in
• One inbox, so capture never costs more than ten seconds
• A weekly processing habit that keeps the system alive without eating writing time
• A citation trail on every note, so the fact-check does not take a season
• Chapter packs that make a writing day start with writing
• A defensible answer to "when is the research done"

The one rule that decides whether it works

Organize by question, not by source.

The unit of book research is not the PDF, the interview, or the notebook. It is the claim: one fact, one quote, one scene, one idea, small enough to sit in a single note. Claims get filed under the question they help answer, and the source they came from travels with them as metadata rather than as the filing system. This inversion is the entire system. Everything else in this guide is mechanics for making it cheap enough that you actually do it.

The test of your current system takes ten seconds: can you pull up everything you know about one of your book's central questions in one view? If the answer involves opening folders, you are organized by source.

How AI fits into book research

Genuinely useful: transcribing interviews, summarizing a sixty-page PDF so you can decide whether it deserves a real read, extracting the claims from a source you have already read, finding the connection you half-remember ("where did I have something about the 1987 audit?"), and clustering a hundred loose notes into themes when you are building boards. Mechanical compression and retrieval of material you gathered.

The poison: letting AI-generated "facts" enter your notes unlabeled. A model asked "what happened in the 1987 audit" will answer fluently whether or not it knows, and once that answer sits in your notes wearing the same clothes as your archival material, your book has a fabrication in it with your name on the cover. The rule: AI output about the world goes in labeled as UNVERIFIED, with no exceptions, and graduates only when a primary source confirms it. AI summaries of your own sources are safe; AI knowledge of your subject is radioactive until checked.

The tooling constraint: research assistance is only as good as what the assistant can see. An AI that reads your actual boards can answer "what am I missing about the flood chapter" from your material; a chat window answers from the internet's material, which is a different book.

1. Define the spine questions before you organize anything

A book is not about a topic; it is the pursuit of a handful of questions through a subject. Before touching the pile, write the three to seven questions the book exists to answer. For the flood book: "Why was the second flood worse when the town was warned?" "Who profited?" "What did the survivors' accounts agree and disagree on?"

These spine questions become your top-level structure: one board or workspace region per question, plus exactly two extras: an inbox, and a parking lot called Not This Book for the fascinating material that belongs to a different project (this drawer is what lets you actually cut things).

Two properties matter. The questions are arguable, like a thesis, not neutral like a category, because arguable questions tell you what evidence is missing. And they are allowed to change: books discover their real questions midway, and when that happens you rename a board, which takes a minute, rather than refiling a folder tree, which takes a lost month. If you cannot write the spine questions yet, that is not a filing problem, and no system fixes it; write the one-page book pitch first.

2. Run one inbox, and make capture cost ten seconds

Every piece of research enters through a single door: one inbox, whatever the material is. A quote from a book, a photo of an archive document, a screenshot, a link, a voice memo from a walk, the transcript of an interview. Capture must cost less than the thought is worth: grab it, throw it in the inbox, keep reading. No filing at capture time, ever, because filing-at-capture is the friction that convinces you to "just remember it" instead, and unrecorded thoughts are where books leak.

The inbox is allowed to be ugly. It is not allowed to be plural. The 212 bookmarks, the four notebooks, and the email attachments were nine inboxes wearing disguises, and nine inboxes means zero inboxes: nothing is anywhere, so everything must be searched for everywhere.

Physical material gets photographed into the inbox; the paper original goes in one labeled box in the order it was photographed. You are not building a paper filing system, you are building a pointer to one.

3. Process the inbox weekly into claims on question boards

Once a week, ninety minutes, non-negotiable: empty the inbox. For each item, break it into claims, one fact, quote, or scene per note, and file each claim onto the spine-question board it serves. Three rules for the pass:

• A claim can serve two questions. Duplicate it. Notes are free; the crime is a claim you can only find by remembering where you put it.
• Write the claim in your own words, with the source's exact words quoted inside it when the phrasing matters. The rewrite is not overhead; it is the moment you find out whether you understood the material, and it is when the claim becomes usable in your voice later.
• Most items produce one to three claims. A major source, a key interview or the central archive file, produces dozens, and that is a sign it deserved the time.

Anything that does not serve a spine question goes to Not This Book or gets deleted. Deleting research feels like burning money and is actually the system working: the questions are doing their job of telling you what the book does not need.

The weekly cadence matters more than the day you pick. Skip two weeks and the inbox becomes a pile; a pile becomes a dread; dread becomes ideas-MASTER-v3.

4. Capture provenance on every claim, at filing time

Every claim carries three fields, filled in during the weekly pass while you still remember: where it came from (source, page or timestamp), what kind of support it is (primary document, participant interview, secondary book, press account, my own inference), and how solid it is (verified, single-source, disputed, UNVERIFIED-AI).

This feels bureaucratic for exactly as long as it takes to reach the fact-checking stage, at which point it is the difference between a week and a season. Nonfiction lives and dies on "how do you know that", and the honest answer must be attached to the claim, not reconstructible in principle from a folder somewhere. My friend's original system had forty-one claims she wanted for chapter one and could source only twenty-eight of; the other thirteen took three weeks to re-find. Filed provenance is three weeks bought for three seconds per note.

The support-type field also quietly runs your remaining research: a chapter resting on press accounts and single sources is a chapter with holes in it, visible at a glance from the board.

5. Lay the material out spatially and let the outline emerge

Here is where organizing by question pays its dividend. With claims on boards, arrange them: chronology across, themes down, contradictions deliberately placed next to each other. Put the two survivor accounts that disagree side by side. Cluster the money trail. Look at the shape.

The outline is discovered in this arrangement, not imposed before it. Chapters are usually visible as clusters that keep asserting themselves: a stretch of chronology dense with claims, a character who owns a region of the board, a contradiction juicy enough to carry thirty pages. This is the step a folder tree can never do for you, because folders cannot show two things next to each other, and next-to-each-other is where books are found. It is also the stage where an AI that reads the board earns its keep: asking for proposed clusters of two hundred claims gives you an arrangement to argue with, and arguing with a wrong clustering sharpens the chapters faster than staring at a list.

Expect this layout to be redone two or three times across the project. That is not churn; that is the book thinking.

6. Build chapter packs so writing days start with writing

When a chapter's cluster looks dense enough to draft, build its pack: one board holding everything the chapter needs. The claims in rough narrative order, the two or three scenes it opens and closes on, the key quotes verbatim with their provenance, the unresolved questions listed at the top, and the handful of claims from other boards it borrows.

The pack is the bridge between the research system and the manuscript, and it exists so a writing morning begins with the material already assembled instead of with a search. Draft from the pack, not from the full research system; the full system is where you go when the pack turns out to be missing something, and each trip back is recorded by adding the found claim to the pack, so the pack stays the chapter's complete record.

When the draft of the chapter is done, mark the claims it used. This is what makes the eventual endnotes and fact-check mechanical instead of forensic: the chapter knows what it is built from.

7. Freeze versions and define done

Two closing disciplines, both about endings, which research otherwise does not have:

• Freeze the research state when a draft milestone lands. When the full first draft exists, the boards get a version marker and stop changing except through the inbox like everything else. Late-arriving material is filed and visibly flagged for the revision pass rather than silently mixed in, so you always know which draft saw which evidence.
• Define done per question, not for the research. Research as an activity never finishes; it is too pleasant, and it is the most respectable available form of not-writing. A spine question is done when its board can answer it with sourced claims and the remaining unknowns are listed and accepted. When all spine questions are done or consciously accepted-incomplete, the research is done, whatever the unread-PDF count says.

Common mistakes

• Organizing by source. The folder of PDFs, the notebook per archive trip. The book asks questions; file by question and let sources be metadata.
• Nine inboxes. Bookmarks, notebooks, email, screenshots folder. One door, or nothing has a place.
• Filing at capture time. The friction convinces you to stop capturing. Capture dumb, file weekly.
• No provenance until the fact-check. Re-finding sources costs weeks exactly when the book is nearly done and morale is thinnest.
• The outline imposed before the material speaks. Chapters invented in advance and stuffed with whatever fits. Lay the claims out and read the clusters instead.
• Unlabeled AI claims in your notes. The one unforgivable one: a fabrication with a book contract. UNVERIFIED until a primary source says otherwise.
• Research as procrastination. Another archive, another PDF, another year. Define done per question and write toward it.
• Keeping everything. Not This Book is where fascinating irrelevance goes to stop hurting you.

How this went for the flood book

Rebuilding my friend's three years of material took two weekends, which surprised both of us; the archaeology dig she dreaded was mostly a filing problem.

Weekend one: we wrote five spine questions after an hour of arguing, made five boards plus Inbox and Not This Book, and processed the 61 PDFs and four notebooks. Most PDFs yielded two to four claims; nine yielded nothing for any spine question and went to Not This Book, which she found genuinely upsetting and then liberating. The interview transcripts were the gold mine: the one with the county engineer produced 31 claims across three boards.

Weekend two: provenance fields on everything (the three-week hunt for the thirteen unsourced chapter-one claims happened here, and converted her to the system more effectively than any argument), then the first spatial layout of the "why was the second flood worse" board. Two survivor accounts that disagreed about the sirens ended up pinned side by side, and she stood in front of them and said "that contradiction is chapter one", which it became.

The system then ran on the weekly ninety-minute pass. Chapter packs got built one at a time as clusters ripened; the first draft took eleven months, and the endnotes, the part she had privately expected to be a second book's worth of suffering, took nine days, because every claim already knew where it came from.

The tools you will actually use

The system is tool-agnostic; the question is where its two core moves, filing claims under questions and arranging them spatially, are cheap.

• Scrivener is the strongest pure drafting environment, and its research folder keeps material near the manuscript. Claims-on-boards is not its shape: material lives as documents in a tree, so the spatial layout step happens somewhere else or not at all.
• Notion and Obsidian both handle claim-notes with metadata well: Notion as databases with source fields, Obsidian as linked markdown with the strongest local-file longevity story. Both are list-and-link shaped: the emergent-outline step, contradictions pinned side by side, clusters asserting themselves, is where they run out of native vocabulary.
• Zotero is the right reference manager for heavy citation loads and pairs well with any of these: it owns the bibliography while your system owns the claims.
• An infinite canvas built for this is where Storyflow fits: question boards as canvases, ten-second capture into an inbox board, claims as cards carrying provenance, and the spatial layout as the native gesture rather than an export. The AI reads the actual boards, so "cluster these two hundred claims" and "what is the flood chapter missing" work on your material, and transcripts or PDFs can be summarized into claim drafts next to the source. It is free to start with unlimited boards and collaborators, paid plans from 7.99 dollars a month billed annually, per account rather than per seat, which matters when a co-author or fact-checker joins.
• The honest summary: the drafting itself is often better in Scrivener or your word processor, and a thousand-source academic bibliography wants Zotero regardless. The case for Storyflow is the middle of the system, claims, questions, and the emergent outline, which is the part folders and lists structurally cannot do.

You are ready

The pile is not the problem, and more discipline was never the answer. The structure was: questions on top, one door in, a weekly pass that turns sources into claims, provenance riding along, and a canvas where the book can show you its own shape.

Write the spine questions. Make the boards. Empty the inbox weekly and delete what serves a different book. Then stand in front of the material and read the clusters, because somewhere on that board, two notes are already sitting next to each other saying: this is chapter one.

Author

Justkay is a documentary filmmaker and the founder of Storyflow. He organizes film research with the system in this guide, has rebuilt more than one author's research pile with it, and built Storyflow so the claims, the questions, and the emerging outline could finally live on the same surface.