Chunk 1,500–3,000 tokens (~1–3 pages) with 100–200 token overlap; chunk by logical section headings. Prompt chain: (1) extract headings, claims and data points with page refs per chunk; (2) produce a 3-bullet micro-summary + one supporting quote; (3) merge summaries into a thematic synthesis and flag contradictions; (4) output a 200–350 word executive brief plus an evidence table. QA: require exact quotes+page numbers, run a claim-to-source checker and re-run low-confidence claims, sample 10% manual checks. For long-context work, consider Claude. Compare options: Compare Claude and ChatGPT