From Notes to Book: Turning Raw Research into Readable Text
Research Notes Are Not Book Text
Notes are written for you; text is written for a reader. A note is compressed, full of private shorthand, and leans on your memory: "go back to the second interview". A book has no such crutch; the reader has neither your memory nor your intent.
Three gaps follow, and all three are yours to close. Context you carry in your head: your note says "compare with the third case" without recording what the third case was. The argument: notes are conclusions, not a route, and the reader wants the route. Orientation: the sentences that tell a reader where they stand have no function in a note, which is why they were never written. Paste notes in sequence and you get a wall of fact — the reader absorbs information and still cannot say why any of it is here. And a note is not a rough draft: a draft is where you decide what to leave out.
Three Kinds of Notes and Where Each One Goes
Before anything else, sort. Sort by the role a note can play in the finished text, not by topic.
| Kind | Tell | Destination |
|---|---|---|
| Evidence and data | Direct quotation, figure, interview excerpt, a line from a source | Enters the text with a source and a clear role; the rest goes to an appendix |
| Thought | Your own conclusion, a hunch, an objection, a distinction you discovered | Claim sentences — the spine of the book comes out of these |
| Process | An open task, an unanswered question, "check this" | Your to-do list; it never enters the book |
The three piles are wildly uneven. A note collection can run to a thousand lines and hold only forty "thought" entries; the book is built from those forty, and everything else supports them.
A one-line test for every note: if this disappears, what breaks in the text? No answer means it is a process note and belongs on the to-do list. If the answer is "nothing, but it is interesting", it is evidence: its home is an appendix or a footnote, not the running text.
Building Structure Out of Notes
Cluster by claim, not by topic. A thematic cluster always looks tidy and always produces a claimless chapter — one that says everything about a subject and proves nothing.
Write one sentence per cluster starting with "This chapter shows that…". If the sentence collapses into "it is about…", the cluster is not a chapter yet. Then chain the clusters: each one comes after the cluster that laid its groundwork. What you end up with is the skeleton of the book.
An orphan note — one that attaches to no cluster — is a signal, not rubbish. It means either a chapter is missing or that note belongs to a different book.
AI as a Cartographer, Not an Architect
Give the model an index of your notes — title and first line of each — and ask for possible clusters, with two constraints stated in the same request: do not invent any note that is not in this list, and name the ID of every member note in each cluster. Without the second constraint you get a tidy cluster whose members do not exist in your index. Then verify by hand; models tend to smooth titles into a neat arc that is not in your raw material. The general path from existing text to chapter headings and final order is in turning a manuscript into a book; the difference here is that the raw material is unpublished and the hard knot is attribution.
From Data to Narrative
A table of numbers is not a finding. Three moves turn data into narrative: choose the contrast, put the strongest evidence in the running text, and send the rest to an appendix. The move everyone skips is the last one: say what the data does not show.
A concrete case: a study with twelve questions. A chapter that charts all twelve leaves nothing behind in the reader's mind; a chapter that picks one contrast and sends the other eleven to an appendix produces one durable finding. Appendices exist for credibility, not for reading.
The model turns a cluster of notes into candidate paragraphs and shows you a sensible order for the data. It cannot know which number will surprise your reader, and worse, it fills gaps. One rule: write the numbers yourself. A model will plug a missing figure with a plausible one, and the sentence will read so smoothly that nobody checks. If part of your raw material was already published, a different logic applies, covered in from blog archive to book.
Citation and Sources: The Editorial Work AI Does Not Do
Accept one plain fact at the start: the model has no automatic source verification and no web search. Page numbers, year of publication, and — most importantly — whether the source actually says this are beyond it.
| Task | Within reach | Out of reach |
|---|---|---|
| Making citation forms consistent | Yes | — |
| Spotting incomplete or unlabelled citations | Yes | — |
| Ordering a reference list to one style | Yes | — |
| Confirming that a source says this | — | No |
| Finding a page number or access date | — | No |
| Telling a direct quote from a paraphrase | — | No; that needs the source itself |
The method: give every evidence note an ID — n-014 will do — and write the source, page, and access date beside it. The ID travels with the note to the last pass. Any citation a model generates is unverified until you open the source yourself; a citation is the ideal home for a plausible error — a complete, tidy format for something that does not exist. The same logic governs quotation: anything inside quotation marks must come from the source word for word, or it is not a quote and must lose the marks.
The Order of Work in Practice
The big picture of editing levels is in the complete map of editing with AI, and I will not repeat it here. What follows is the working order for one specific input: a pile of notes.
- Sort the notes into evidence, thought, and process.
- Write one claim sentence for every "thought" cluster, starting with "This chapter shows that…".
- Order the clusters by dependency; that becomes the skeleton and the final running order.
- Draft each chapter rough: put the evidence notes in place and do not polish sentences.
- List the gaps — where a claim has no evidence — and go after those gaps instead of rereading the whole archive.
- At the structural level, move, merge, cut; keep the cut material in a file rather than deleting it.
- Attach and standardize citations, then build the reference list; chase down every citation without an ID.
- Only now does the sentence come first: line editing, and copyediting after that.
Keep two files open and current for the whole project: the gap list and the unverified-sources list. If you cannot open a source by the end, its sentence does not go in the book. That is the only pass criterion that matters.
