How to Feed Claude 3.5 Sonnet a 100-Page Document
Claude 3.5 Sonnet’s long context is the feature that made it popular with analysts. It can read a 100-page PDF in one go — but ‘can’ and ‘should’ differ. Feed it everything at once and the model gets vague, because the signal drowns in the noise.
Better results come from structure: a clear question, the right slice of the document, and an explicit output format.
How Much Can Claude 3.5 Sonnet Read at Once?
The 3.5 Sonnet line supports a context window in the 200K-token range, enough for a few hundred pages of text. That is a ceiling, not a target. Accuracy on a specific fact degrades as you add unrelated pages, and cost climbs with every token you send.
If your question is about chapter 4, send chapter 4.
How Do You Structure a Long-Document Prompt?
- Put the document first, the question last — the model reads in order.
- Scope it: “Answer only from sections 3 and 4.”
- Request a shape: a table with columns Claim, Source Section, Confidence.
- Ask it to quote the line it used for each claim.
The quote requirement is the single biggest accuracy jump. A model that must cite a line is far less likely to invent one.
What Are Common Long-Context Mistakes?
- Dumping the whole corpus when one section answers the question.
- No output schema, so answers arrive as loose prose.
- Trusting a summary without checking the quoted source.
- Reusing one giant thread across unrelated tasks, which muddies context.
When Should You Split Into Chunks Instead?
For very long corpora or repeated queries over one document, chunking pays off: split by section, run each through Claude 3.5 Sonnet, then ask it to reconcile the summaries. Costs less, and each answer stays grounded. For a single focused question about a mid-size report, one well-scoped prompt beats a pipeline.
Key Takeaways
- 3.5 Sonnet supports ~200K tokens of context — a ceiling, not a goal.
- Send the relevant section, not the whole document, for sharper answers.
- Require quoted source lines; it cuts hallucination on long texts.
- Chunk by section for repeated queries over large corpora.
