How to analyse a whole book with ChatGPT

Why one upload does not work

Two limits, and the second is the one people underestimate.

The token cap. Text and document files are capped at 2 million tokens. A long book can approach this, and past it the file is truncated with no warning at all.

Attention. Well before any hard limit, quality falls. Ask a question about chapter nine having sent the whole book, and the answer draws heavily on whatever was most prominent, which is usually the beginning and the end.

So even when a book technically fits, sending it whole gets you a worse read than working through it.

The approach that works

1. Convert to plain text and split by chapter

   Chapters are natural boundaries and they match how you will ask questions. Markdown is ideal, since the headings survive.

2. Set the frame once, at the start

   "I am working through a book chapter by chapter. After each chapter, give me: the main argument in three sentences, any claim that needs evidence, and how it connects to earlier chapters. Do not summarise the whole book."

3. Send one chapter, get the notes, keep them

   Copy the output into your own document. This becomes your running record, and it is what makes the rest work.

4. Re-paste the accumulated notes every few chapters

   "Here are my notes on chapters 1 to 6. Chapter 7 follows."

   This is the step that matters. It puts the earlier material back in view rather than trusting the conversation to have kept it.

5. Ask whole-book questions at the end, against the notes

   Not against the book, which is no longer in view, but against the notes you have built. They are a condensed version of it and they fit comfortably.

Checking each chapter landed

Long chapters get truncated on their own. Ask something only answerable from the end of the chapter:

What does the final paragraph of this chapter say?

Ten seconds, and it catches the failure that otherwise poisons everything downstream.

Questions that work well on a book

Per chapter: what is the argument, what evidence supports it, what is asserted without support.

Across chapters: where does the author contradict something said earlier, and which claims does the book depend on but never establish.

At the end, against your notes: what would a well informed sceptic say, what has been left out, and what would you have to believe for the conclusion to hold.

That last set is where this becomes worth the effort. Anyone can get a summary. Finding the argument that does not hold together across three hundred pages is genuinely hard, and it is a question the notes can answer.

When to use something else

Being straight about it. If you are doing this often, chunking a book is the wrong tool.

Assistants with much larger context windows can hold a book in one piece, which removes the whole problem. Retrieval tools that search a document and surface the relevant passages are built for exactly this and will beat any chunking strategy on questions like "where does the author discuss X".

One evening on one book, the approach above is fine. A shelf of books every month, and it is worth changing tools.

Scanned books

OCR first, without exception. Every plan except Enterprise discards images, and a scan is entirely images, so there is nothing to read until you create text.

Expect errors on older print, and skim the OCR output before working from it. Mangled words produce confidently wrong answers rather than obviously broken ones, which is the harder failure to notice.