How to Brief Claude for Long Documents (SOPs, Policies, Guides)
Claude is strong on long, structured docs β if you brief it like a junior hire, not like a magic box.
I learned this writing SOPs: vague brief β confident garbage. Tight brief β usable first draft.
Brief template
Document type: [SOP / policy / internal guide]
Audience: [who will read it]
Must include: [sections]
Must avoid: [legal claims / tools we donβt use]
Tone: plain English, steps a new hire can follow
Constraints: Max [X] words per section. Numbered steps where possible.
Source notes: [paste messy notes]
Task: Draft the full document structure + content.
QC checklist
- Any step you couldnβt follow cold? Rewrite
- Any tool/version invented? Remove
- Any policy language too strong? Soften or lawyer-review
AI drafts process docs well. Humans own accuracy.
Taming the 200k Context Window
Anthropic's Claude is famous for its massive context window. You can upload entire books, employee handbooks, or massive codebases into it. However, if you just dump a 50-page PDF into Claude and say "summarize this," it will miss the critical nuances. To get professional-grade results, you have to structure your input data.
Using XML Tags for Structure
Claude is explicitly trained to pay extreme attention to XML formatting. If you want it to process a long document perfectly, wrap your document and your rules in tags. For example:
[Paste your 30-page SOP here]
</document>
<rules>
1. Extract all mandatory compliance steps.
2. Ignore the historical preamble.
</rules>
The "Chain of Density" Summary
When briefing Claude on long documents, I use a technique called the Chain of Density. I prompt it: "Read the document. Write a 5-sentence summary. Then, rewrite the summary to be 50% shorter but pack in twice as many specific entities (names, dates, metrics). Do this 3 times, giving me the final, incredibly dense version." This forces the AI to cut the fluff and deliver only the hard data you actually need.
β Muhammad Bilal Khan, Founder at PromptHub.