How to Organize Research for a Book (2026 Step-by-Step Guide)

Key Takeaways

• Organize by destination, not by subject. A note tagged "Chapter 4, the objection about cost" is usable. A note tagged "economics" is a library card.
• Separate capture from processing. Capture fast and messy into one inbox. Process on a schedule. Trying to file perfectly at capture time kills the capture habit.
• Keep one source of record. Full citation, page number, and a direct quote at capture time, because going back to find a page number eighteen months later is the most demoralizing work in publishing.
• Distinguish the three kinds of research note: evidence you will cite, context you need to understand but will not cite, and texture that makes scenes real. They have different lifespans and different homes.
• The outline is the filing system. You cannot organize by destination until a provisional structure exists, so write a rough outline early and expect it to change.
• Most research never appears in the book, and that is correct. The usual ratio is somewhere between five and ten to one. Collecting more than you use is the job, not a failure.
• Review the whole set before drafting. Laying the evidence out against the outline is where you find the chapter with no support and the chapter you over-researched.
• AI is useful for summarizing and gap-finding, and unreliable for facts. Never let a model be your source of record.

The Library and the Draft

There are two ways to organize research, and they optimize for opposite things.

The Library organizes by subject. Everything about the labor movement goes under "labor," everything about a person goes under their name. It is the intuitive system because it mirrors how libraries and encyclopedias work, and it optimizes for one question: *where do I put this?*

The Draft organizes by destination. Everything that will support the argument in chapter four goes with chapter four, regardless of subject. It optimizes for a different question: *what do I need in front of me to write the next two thousand words?*

The Library question is the one you face at capture time, hundreds of times. The Draft question is the one you face at writing time, and it is the one that actually blocks books. Which is why building only a Library feels productive right up until the day you sit down to write and discover your notes cannot tell you what to say.

The honest complication is that you need both, and in that order. Early on you genuinely do not know where anything goes, so you build a Library by necessity. The mistake is staying there. The moment you have a provisional outline, the filing system should change from subject to destination, and most authors never make that switch because nothing forces them to.

The Method, Step by Step

1. Set up one inbox, not five

Every piece of research enters through a single door: a note in one app, a folder, a physical tray for paper, whatever you will actually use. The requirement is singularity. Research scattered across browser bookmarks, three notes apps, camera-roll photographs of book pages, and email-to-self is not a research system, it is four systems that each contain a fifth of the answer.

Capture should be nearly frictionless. If capturing takes more than a few seconds, you will stop doing it while reading, which is exactly when the good material appears.

2. Capture the source of record at the same moment

This is the one piece of discipline that pays for itself repeatedly. At capture time, record: the full citation, the page or timestamp, and a verbatim quote of the passage that matters, marked clearly as a quote.

The reason is brutally practical. Eighteen months later, in copyedit, someone will ask for the page number of a claim on page 212, and if you did not capture it then, you will spend a full day of your life finding it. Authors who skip this step pay for it in a single miserable week near the end.

Distinguish, visibly, between a verbatim quote, a paraphrase, and your own thought. Notes that blur these three are how accidental plagiarism happens, and it happens to careful people who took a fast note at midnight and could not tell, a year later, whose sentence it was.

3. Write a provisional outline earlier than feels comfortable

You cannot file by destination without destinations. The outline does not need to be right, it needs to exist: a list of chapters with a one-line statement of what each one argues or covers.

Expect it to change substantially. That is fine, and refiling is cheap compared to drafting from an unusable pile.

4. Process the inbox on a schedule

Set a recurring session (weekly works for most) where you empty the inbox. For each item, make three decisions:

• Type. Is this evidence you will cite, context you need but will not cite, or texture (a detail, an image, a scene-setting fact)?
• Destination. Which chapter, and which section within it? If it does not have a destination, that is information: either the outline is missing something or the note is not for this book.
• Function. In one line, why does this belong? "Supports the cost objection." "Contradicts the standard account of the 1974 decision." This line is what makes the note usable a year later.

That third decision is the highest-value ten seconds in the entire process. A note that says what job it does can be acted on. A note that is just a highlighted passage requires you to re-read and re-decide, which is the tax that makes writing from research feel so slow.

5. Keep a "does not fit" pile, and keep it visible

Material that contradicts your argument or refuses to fit anywhere is the most valuable research you have, and the easiest to quietly lose. A contradiction is either the thing that makes the book honest or the thing a reviewer will use against it. Keep them together, deliberately, and force yourself to address them.

6. Lay the whole thing out before you draft

Before writing begins, get all of it visible at once: chapters across one axis, evidence under each. This is the single most useful hour in the process and the one most authors skip, because it requires a surface bigger than a document.

What you are looking for is imbalance. A chapter with three thin sources and a chapter with forty. A claim in the outline with no evidence under it at all. An argument that is entirely supported by one source, which is a structural risk rather than a research one.

This is the moment a visual canvas earns its place. A document shows one screen at a time, and imbalance is invisible one screen at a time. Storyflow is useful here specifically because its AI reads the whole board you are working on, plus up to one Tactic and three Documents you `@`-mention, so you can ask which chapter is thinnest while the actual evidence is on the board being read, rather than pasting a summary into a chat that cannot see any of it.

Where Storyflow is the wrong tool: it is not a citation manager, so Zotero or a comparable reference tool still owns bibliography, formatting, and page-level citation. It is not where the manuscript gets written, so Scrivener, Ulysses, or Word still owns the draft. It is cloud-only, which rules it out for authors working with confidential or embargoed material that must stay local. And an author working purely from printed sources, in a linear argumentative book, may find a canvas to be overhead that a chapter-folder structure handles fine.

7. Re-file as the outline changes

The outline will change, and when a chapter splits or dies, its research needs to move with it. Authors who skip this end up with a filing system describing a book they are no longer writing, which is worse than no filing system because it is confidently wrong.

The Three Kinds of Research Note

Treating all research as one kind of thing is a common and expensive mistake. There are three, and they behave differently.

Evidence is material you will cite: statistics, quotes, primary documents, study findings. It needs the strictest handling, because every piece will eventually be checked. Full source, page, verbatim quote, no exceptions.

Context is what you need to understand the subject well enough to write about it confidently, most of which will never appear in the book. Its job is to be absorbed, not filed. Over-organizing context is a common time sink, since you are building retrieval infrastructure for material you will never retrieve.

Texture is the specific detail that makes writing feel real: what the room smelled like, what someone was wearing, the make of the car. Texture is easy to lose because it looks trivial at capture time, and it is disproportionately what readers remember.

Filing these three identically means either over-processing your context or under-capturing your evidence. Most authors do both.

The Note System Underneath the Tool

Before the software question, there is a structural one that most authors answer by accident: how are notes filed relative to the book? There are three answers in common use and they behave very differently at scale.

Chapter buckets. Every note goes under the chapter it serves. Simple, fast to use, and it is what most first-time authors build. The failure arrives at the first restructure: when chapter four dissolves into chapters three and seven, its notes have to be re-sorted by hand, and any note that served two chapters was already a problem. Works well for books with a fixed, externally determined structure (a biography running chronologically, a manual following a process) and badly for arguments still being worked out.

Thematic notes, linked. Notes are filed by idea rather than by destination, and connected to each other. This is the Zettelkasten family, and its advantage is real: because notes are not committed to a chapter, restructuring costs nothing, and connections between ideas surface that a chapter filing would have hidden. Its failure mode is equally real, and it is that the system becomes the hobby. An author with 900 beautifully linked notes and no chapters has built a library, not a book, and the linking gives that state a satisfying feeling of progress.

Commonplace, chronological. One running document, everything appended in the order you found it, searched rather than filed. Sounds primitive and works better than expected for books under a certain evidence load, because it has no maintenance cost at all and search is very good now. It breaks when you need to see everything on one theme at once and the answer is spread across 200 pages of chronology.

The practical recommendation for most books: thematic notes during collection, chapter buckets from the outline onward, and a physical or visual layout at the transition between them. The transition is the moment step six describes. Notes stay unfiled while the structure is genuinely open, then get committed once it is not, and the layout is how you decide it is not.

Choose deliberately, because the cost of switching systems mid-book is high enough that most authors do not, and instead spend the second half of the project fighting a decision they made in week one without noticing they made it.

Research Tools Authors Actually Use

Most authors end up with two or three of these, not one: a citation manager, a place notes live, and a place the manuscript gets written. The tools below are grouped by which of those jobs they do, because a tool evaluated for the wrong job always looks bad.

Scrivener

Still the default for book-length work, and the reason is structural rather than featural: the research folder sits inside the same project as the manuscript, so a source is two clicks from the paragraph it supports. The corkboard and outliner make restructuring genuinely cheap, which matters more than any other property in a long project, because restructuring is what long projects do.

The cost: a steep learning curve that is not exaggerated, and a compile step that has defeated experienced writers. It is also weak at capture, so most Scrivener users pair it with something faster for the collection phase.

Pick it when: the book is long, the structure will move, and you are willing to spend a weekend learning a tool you will use for two years.

Zotero

The standard for citation management and free, which makes the decision easy. It captures sources from the browser with one click, stores the PDFs, and generates formatted citations in any style a publisher demands. If your book has a bibliography, this or an equivalent is not optional, and the authors who decide it is optional discover the cost during fact-check with a deadline.

The cost: almost none. Storage above the free tier is paid, and the interface is functional rather than pleasant.

Pick it when: you cite anything. This is the least debatable recommendation on the list.

Obsidian

Local-first plain text with links between notes, and the natural home for the thematic system described above. Files are yours, on your disk, in a format that will still open in thirty years, which is a genuine consideration for a project measured in years. The plugin ecosystem runs past 5,000 and covers most of what the core app does not.

The cost: it is an excellent Library and becomes a Draft only through deliberate effort, because nothing in the tool pushes you toward a finished linear artifact. This is the tool where the note system most often becomes the hobby.

Pick it when: you want to own your files and you have the discipline to outline anyway.

Notion

A table of sources with status, chapter, type, and "used yet" columns is a legitimately good processing system, and for authors who think in databases nothing else feels as orderly. Filtered views by chapter are the specific feature that earns its place here.

The cost: slower to capture into than a dedicated notes app, and everything lives in someone else's cloud. AI is bundled into the Business tier at $20 per user per month since the standalone add-on was retired in May 2025, which is a lot for a solo author.

Pick it when: your research has structured attributes you want to filter and sort.

DEVONthink

The heavyweight for research-dense nonfiction, and the tool historians and investigative writers reach for when the corpus runs to thousands of documents. Its search and its "see also" suggestions across a large collection genuinely surface things you did not remember having.

The cost: macOS-focused, priced as professional software, and considerably more tool than a book with 60 sources needs.

Pick it when: the archive is large enough that finding things is the bottleneck.

Storyflow

Our product, and it earns one job in this workflow: the step-six review where the whole structure has to be visible at once. Chapters as columns, evidence under each, and the gaps immediately legible because a column with two cards under it looks different from a column with thirty. Its AI reads the whole board plus one Tactic and up to three @-mentioned documents, so with the outline and the evidence on the same board you can ask which chapters are thin, or which two chapters are making the same argument.

Where it is the wrong tool: it is not a citation manager, so Zotero still does that job and does it better than anything bundled would. It is not where a manuscript gets written, so the draft lives in Scrivener, Word, or Google Docs. It is cloud-only, which rules it out for research that cannot leave your machine. And for a book with 40 sources, index cards on a table do the same job for free. Free plan available; Plus is $7.99 a month billed annually or $9.99 monthly.

Pick it when: the evidence is gathered, the outline is provisional, and you cannot see the whole thing.

Milanote

The other honest answer for the layout stage, with a longer track record and a calmer surface. No AI reading the board, which is the only meaningful difference for this job, and for many authors that is a feature rather than a lack.

Index cards and a large table

Not a joke and not nostalgia. For a book with a manageable evidence base, physical cards laid out on a table beat every tool on this list at exactly one thing, which happens to be the important thing: peripheral vision. You see the whole structure without scrolling, you rearrange with both hands, and physically moving a card is a decision your body registers. The limits are obvious (no search, no backup, no version history) and they are the reason the digital options exist.

Where AI Helps and Where It Will Hurt You

AI is genuinely useful in two places in this process, and dangerous in a third.

Useful for summarizing volume. Given a long report or transcript, a model will produce a serviceable summary that helps you decide whether the source deserves a full read. This is triage, and it saves real time.

Useful for gap-finding. Given an outline and the evidence filed under it, a model is reliably good at noticing that chapter six has no support, that two chapters make the same argument, or that a claim rests on a single source. Structural absences are exactly what models notice well.

Dangerous as a source. Models produce confident, well-formed, entirely fabricated citations, and they do it most often for the kind of plausible-sounding reference an author would not think to check. Never let a model be your source of record. Every fact and citation in a book must trace back to something you personally opened.

The line worth holding: use AI on your research, never instead of it.

Organizing Research for a Novel

Everything above assumes nonfiction, where research supports an argument and every claim needs a source behind it. Fiction research works differently enough that applying the nonfiction system to a novel produces a very organized book that nobody wants to read.

The purpose is different. Nonfiction research proves. Fiction research furnishes. You are not building a case, you are building a world dense enough that the reader believes the parts you invented. That changes what is worth keeping: the specific detail beats the comprehensive account every time. Knowing that a 1950s Glasgow tenement had a shared toilet on the half-landing is worth more than a chapter of housing policy, and the housing policy is what your notes will fill up with if you file by topic.

File by story element, not by subject. Character, place, object, period texture, and procedure are the useful buckets, because those are the units a scene is made of. A subject-filed note ("Victorian mourning customs") sits inert. The same material filed under a character ("what Ellen does with her sister's ring") is already a scene.

Keep a strict separation between research and invention. Two colours, two folders, whatever you like, but keep it absolute. The failure mode is specific and common: eighteen months in, you cannot remember whether the detail about the harbour bell is something you read or something you made up, and now you either cut a good detail or risk asserting something false about a real place. Mark it at capture. It takes a second and it cannot be reconstructed later.

Cap the research. Fiction research is the most enjoyable form of procrastination available to a novelist, because it is genuinely productive-feeling and has no natural end. Set a date, not a volume. Then write the draft and mark the gaps with a placeholder rather than stopping to look things up, because stopping to look things up is how a writing session becomes a reading session.

The period-texture file is the one worth building carefully. Prices, slang, what things smelled like, what a working day contained, what was on the radio. These are the details that do the believing work, and they are almost impossible to find on demand mid-scene, which is exactly when you need them. A running file of texture, filed by year or era rather than by chapter, pays back more than any other fiction research habit.

The layout step still applies, and applies with more force. Scenes on a canvas with the timeline visible catches the two problems that plague novel structure: a character who disappears for 120 pages, and three consecutive scenes doing the same emotional work in different rooms.

A Worked Example: One Nonfiction Book, 340 Sources

A journalist writes a book on the collapse of a regional bank. Two years, 340 sources: interviews, court filings, internal emails obtained through discovery, regulatory reports, and contemporaneous news coverage.

The inbox. Everything lands in one place, a single Obsidian vault, captured with the source of record attached at the moment of capture. This sounds fussy for two years and it is the reason fact-check took three weeks instead of three months. The rule held throughout: no note enters without its source, and a note whose source cannot be established gets marked as unusable rather than quietly kept.

The provisional outline, written in month four. Eleven chapters, written far earlier than felt comfortable, when perhaps a fifth of the research existed. It was wrong in specifics and right in shape, which is the whole point. Its job was not accuracy. Its job was to give incoming material somewhere to go, so that research stopped being accumulation and started being a search for particular things.

The processing session, weekly, ninety minutes. Each note gets a chapter, a "does not fit" mark, or deletion. The does-not-fit pile stayed visible and grew to about 40 items, of which two eventually became the spine of a chapter that was not in the original outline: the regulator's internal disagreement, which nobody had set out to write about because nobody knew it existed until three filings lined up.

The layout, month sixteen. Chapters as columns, evidence under each, the whole thing visible at once for the first time. Three findings in one afternoon. Chapter seven had 31 sources and chapter nine had four, and chapter nine was the one carrying the central argument. Two chapters were making the same case about incentive structures in different words. And the strongest interview in the whole project, the one everything else was arranged around, was filed under a chapter it did not actually serve.

What the layout was worth. None of those three things were visible in a document, because a document shows you one chapter at a time and the problems were relational. Fixing them before drafting cost an afternoon. Discovering them in the manuscript would have cost a restructure at 90,000 words, which is the specific disaster this whole method exists to avoid.

The honest part. The system did not make the book good. Two years of reporting made the book good. What the system did was prevent the two failures that had nothing to do with the reporting: losing the source for a claim under deadline, and discovering a structural problem after the structure was written into prose.

The Bottom Line

Organizing research for a book is not a filing problem, it is a retrieval problem, and the two want opposite systems. Capture into one inbox, record the source at capture time, write a provisional outline early, then process by destination rather than subject so that every note knows which chapter it serves and why. Lay the whole set out against the outline before you draft, and expect to find one chapter starving and another overfed.

Research organized by subject builds a library. Research organized by chapter builds a book. The switch from the first to the second is the moment a pile of notes becomes something you can write from, and nothing in your tools will prompt you to make it. You have to decide to.