How to split a long document for ChatGPT

The instruction matters more than the cutting

Most advice about splitting explains where to cut and stops there. Getting the cutting slightly wrong costs you a little. Skipping the instruction ruins the whole exercise.

ChatGPT's default is to respond to each message. So without being told otherwise, it reads part one, summarises it, and by the time part six arrives it has been forming and revising an opinion for five messages. The result is anchored to the opening.

Send this before any content:

I am going to send a long document in several parts. Do not summarise, analyse, or answer anything until I tell you the last part has arrived. After each part, reply only with "received".

The "reply only with received" clause earns its place. Without it you get a paragraph of commentary after each part, which is noise and which nudges its thinking early.

Where to cut

At section or chapter boundaries wherever the document has them. A section read across two parts is harder to follow than one read whole.

At natural breaks if there are no headings. Between topics, between arguments, between dates in a log.

Not mid sentence if you can avoid it, though it matters less than people assume. Mid argument is the real cost.

How big each piece should be

Under 10,000 characters, which is roughly 1,500 words.

That ceiling is not a style preference. Paste more than that into the composer and ChatGPT converts the paste into a file attachment rather than leaving it as text, which spends an upload slot on every part and defeats the point of splitting by hand. The threshold was 5,000 characters until June 2026, so older advice sets it lower.

Within that ceiling, aim high rather than low. Parts of eight or nine thousand characters keep the count down, and a coherent section spread across three messages reads worse than one sent whole.

If a paste does turn into an attachment anyway, ChatGPT offers a "Show in text field" control on that individual paste, which puts the text back inline.

Doing it by hand, step by step

1. Convert to plain text first

   Removes formatting and images, and makes the document far easier to cut. It also shrinks the file dramatically if it was image heavy.

2. Find your cut points

   Section headings are ideal. Aim for pieces under 10,000 characters, roughly 1,500 words.

3. Send the instruction

   The wait-for-all-parts message above, before any content.

4. Send each part in order, labelled

   Wait for the "received" before sending the next one, so nothing arrives out of order.

5. Say when you are done, then ask

   "That was the final part. Now: [your question]."

6. Check it got everything

   Ask what the final section is called. If it cannot say, a part went missing.

That last step is not optional. The most common failure of manual splitting is a part that was skipped or that silently truncated, and you will not notice from the answer.

What splitting costs you

Worth being honest about.

Connections between distant sections are weaker than if the whole document had been read at once. Ask about something in chapter two and something in chapter nine together, and the answer is less sure than it would have been.

For most questions this does not show. For questions genuinely about the shape of the whole document, it does, and at that point a tool with a larger capacity is the better answer than more careful chunking.

When not to split

When you only need one section. Send that section. Better answer, less work.

When it is a scan. Splitting images achieves nothing. OCR first.

When it is a spreadsheet. Spreadsheets are exempt from the token cap, so a large one often goes through whole. Splitting a dataset also breaks any analysis across the whole of it.

When it is genuinely book length. Even in order, the earliest parts fall out of memory before the last ones land. A retrieval tool that searches the document will beat any chunking strategy.