Summarise Long Docs in Chunks, Then Combine
What it is
A workflow for documents too long to summarise reliably in one pass: summarise each section on its own, then ask the model to synthesise the section summaries into a single coherent overview.
Why it works
On a very long input, a one-shot summary skims — it over-weights the start, thins out later, and drops detail unevenly. Summarising section by section keeps each part in focus, and combining the pieces produces balanced coverage. You also get intermediate summaries you can spot-check against the source.
When to use it
Long reports, books, transcripts, or multi-document sets where a single summary would inevitably lose the back half.
When not to use it
Documents that comfortably fit the model's attention in one pass — chunking those just adds steps for no gain.
Prompt
This document is long. Step 1: summarise each section separately in <N> bullets, labelled by section. Wait after each. Step 2: once we've done all sections, combine them into one <length> overview that keeps the balance across the whole document.Example
Summarising a 60-page report chapter by chapter, then combining, captures the recommendations in chapter 9 that a single-pass summary quietly skipped.
Advanced version
After combining, ask it to reconcile any contradictions it noticed between sections — long documents often disagree with themselves, and surfacing that is more valuable than a smooth summary.
Common mistakes
- Asking for one summary of a huge doc and losing the later sections.
- Not labelling section summaries, so the combine step blurs where things came from.
- Skipping a spot-check of the section summaries against the source.