How to make ChatGPT follow a document's structure

A generic summary flattens a document. Fourteen sections go in, five paragraphs of themes come out, and the mapping between them is gone. That is fine when you want the gist. It is useless when you need to check a specific clause, brief someone section by section, or compare two versions of the same report.

The fix is to stop asking for a summary and start asking for a structured pass. The structure has to come from the document, and it has to be visible in the conversation before any analysis happens.

Extract the outline before asking anything else

Make the first message do nothing but map the file.

List every heading and subheading in this document, in order, using the document's exact wording. Number them. Do not summarise or comment on anything yet.

Two useful things happen. You get a map you can compare against the real document in a few seconds. And the outline now exists as plain text in the conversation, which is a far more stable reference than the file itself for the rest of the session.

Read the list properly before continuing. If headings you know are in the document are missing, the problem is not your prompt. Anything past 2 million tokens is cut off silently, so a short outline for a long file usually means truncation rather than a badly formatted document.

Hand the outline back as the shape of the answer

This is the instruction that does the work:

Use the numbered headings above as the headings of your answer, in the same order and the same wording. Do not add headings, rename them, reorder them, or merge two into one. If a section has nothing material in it, keep the heading and write "nothing material" underneath.

The last sentence matters more than it looks. Without it, empty sections get quietly dropped, your numbering stops matching the document, and you cannot tell whether section 9 was skipped because it was irrelevant or because it was never read.

The explicit ban on merging matters too. Collapsing adjacent related sections is what summarising normally does, so it will happen by default unless you rule it out.

Work through it section by section

For anything past a few thousand words, one request for the whole document produces detailed early sections and thin late ones. Batching fixes it.

1. Get the outline

   Extract and number the headings, and confirm the list matches the document.

2. Fix the output template

   Decide the fields you want per section and state them once, so you do not restate them every time.

3. Ask for three or four sections at a time

   "Now do sections 1 to 4, using the template." Then 5 to 8, and so on.

4. Check the headings back

   At the end, ask it to list the headings it used. Compare against the original list. Any difference is a section it skipped or invented.

Slower than one prompt, and considerably more accurate. It also means you can stop after the sections you actually care about.

Template the response format

If you want output you can scan, paste, or diff against another document, fix the fields:

Then: "Repeat this template for every section. Keep the field names identical. Do not add fields."

The Figures row is the one people leave out and regret. Requiring the source sentence for every number turns a summary you have to trust into one you can verify without reopening the file.

Keep the structure across a split document

When a document has to go in as several parts, number the sections globally rather than per part. Section 1 to 12 across the whole report, not 1 to 4 three times over.

Before each new part, restate the running outline and say where the new material picks up. At the end, ask for one consolidated pass over the full numbered list. Without that, you get three disconnected analyses whose section numbers collide, which is worse than a generic summary because it looks organised.

Checking it actually followed the structure

Three quick tests, in order of how often they catch something:

• Ask it to list the headings it used, and compare against the outline from step one.
• Pick a section in the middle and ask what the document says under it. Middle sections are where attention thins first.
• Ask what is in the final section. If it cannot say, the file was truncated and everything after that point is missing rather than merged.

What to do next time

Extract the outline. Read it. Hand it back with an explicit ban on adding, renaming, merging and dropping. Fix a template. Work in batches of three or four sections. Check the headings at the end.

That is five extra sentences of prompting, and it is the difference between output shaped like your document and output shaped like every other summary.