UsageAnswered

When summarizing long documents, is there a way to prevent missing the middle content?

Anonymous8/19/2026· Views 88

When I have Claude summarize a PDF report over 100 pages, it seems to summarize mainly the beginning and conclusion, and glosses over the middle chapters. Other than dividing it by chapter and summarizing each separately, is there a prompt method to summarize the entire thing all at once without missing anything? I'm also wondering if I just have to split it up myself anyway.

Answers 1

You need to log in to write an answer.

첫째StaffLv.68/19/2026

Hello, this is the admin speaking. To get straight to the point: simply changing how you structure your prompt can significantly reduce missing middle content, and you don't need to split it up by chapters. When you put a long document in all at once and immediately ask for a summary, there's a tendency to focus only on the beginning and conclusion. This is a phenomenon seen across language models that handle long contexts, and Anthropic's official documentation provides prompt techniques to address this. 1. Place the document at the very top of your prompt, and instructions (like 'summarize this') at the very bottom. If instructions come first, there's a greater tendency to quickly skim through the document in line with the instructions. 2. Divide the document using XML tags like <document>, and chapters with <chapter1>, <chapter2>, etc. When you explicitly structure with tags, each chapter is recognized as an individual unit, which reduces the chance of skipping. 3. Don't ask for a summary right away. Instead, first instruct the model to 'extract key sentences from each chapter as direct quotes,' then have it write a final summary based on those quotes (the quote-then-summarize method). Anthropic's own testing has confirmed that this approach improves recall accuracy on long documents. 4. If you enforce the output format to be a table or list with '2-3 sentences each for chapters 1 through N,' it becomes structurally impossible to skip middle chapters due to format constraints. Even combining these methods, if the document is very long or there's a large difference in content density between chapters, splitting it up as you mentioned might be more reliable. However, in most cases, just using these methods significantly reduces the perceived gaps, so I recommend trying them first. I've put together a detailed guide on this topic below: https://usingclaude.com/ko/library/prompts/long-context-document-analysis-tips