Use hierarchical chunking + a rolling summary:
- Chunk by logical headings (~2–5k tokens) with 200–500 token overlap.
- Prepend a concise running summary to each chunk; update it after processing.
- Build an index of chunk titles + keywords and retrieve only relevant chunks for queries.
- Progressive summarization: extract facts, then synthesize.
- Save chunk embeddings for retrieval (RAG) to avoid re-sending whole doc.
Claude is built for long-context work—set a system instruction to preserve the running summary and token budget.

Compare Claude’s long-context strengths: Compare Claude and ChatGPT