Recommendation
Use a staged Claude pipeline: (1) ingest + normalize PDFs, (2) chunk and summarize each chunk with provenance, (3) build a lightweight index, (4) run one or more synthesis passes that merge chunk summaries into sections, and (5) produce a final cohesive review with a verified reference list. This keeps outputs traceable and enables targeted re-runs when new papers arrive.

Decision criteria (how to choose chunk size and strategy)
- Need high-fidelity quotes and exact page refs: chunk small (≈300–700 tokens; ~half to one page) and keep 20–50 token overlap. Preserve page numbers and paragraph offsets. Higher cost and more calls, but better provenance.
- Need speed and lower cost: chunk larger (≈1,000–2,000 tokens; ~1–2 pages) with minimal overlap. Less accurate quoting but fewer API calls.
- Team size / workflow stage: solo researcher can iterate interactively; larger teams should add a metadata extraction step and canonical citation keys so collaborators can review provenance.

Chunking strategy (practical)
1. Run OCR + extract plain text by page. Keep page number metadata. 2. Split by explicit section headings when present, else split every ~1 page into chunks of 400–800 tokens. 3. Add 100–200 token overlap to preserve context for sentences that cross splits. 4. Assign canonical chunk IDs: PAPERID_pXX_cY.

Prompts (templates)
Chunk summarizer (one call per chunk):
"You are a meticulous researcher. Source: {{PAPERID}}, page {{PAGE}}, chunk {{CHUNKID}}. Produce: (A) 2–4 sentence summary; (B) list of 3–6 bullet takeaways including any quantitative claims; (C) direct quotes (verbatim) up to 200 chars, each with exact page and paragraph or chunk id; (D) extracted citations or DOIs if present in the chunk. Output JSON with fields: chunk_id, summary, takeaways[], quotes[{text, location}], extracted_refs[]."

Mid-level synthesizer (merge N chunk summaries about a topic):
"Given these chunk summaries (include chunk_id and source metadata), produce a single 150–300 word integrated subsection. Keep claims traceable by inserting citation keys like [PAPERID_pXX]. Note disagreements and unresolved open questions. Provide 3 suggested inline citations (canonical keys) and a confidence score (low/med/high) for factual accuracy."

Final review prompt (one pass over section-level syntheses):
"Produce the final literature review section. Merge subsections into a cohesive narrative with transitions, produce a short conclusion and a formatted reference list. For each in-text citation, include a matching reference with authors, year, title, DOI, and original page ranges. Add a verification checklist at the end noting any quotes that need human validation."

Citation handling tips
- Extract and attach full metadata to each chunk (authors, year, title, DOI, pages). Use deterministic citation keys (AuthorYYYY_pXX).
- Keep both inline provenance keys and a final canonical reference list. Always mark verbatim quotes for human verification.
- For accuracy-critical work, run a verification pass that asks Claude to find the quoted sentence in the original chunk and return the exact text and location before you publish.

Best-for / Avoid-if
- Best for: multi-paper thematic synthesis, traceability, iterative review cycles.
- Avoid if: you only need a quick one-paragraph overview (use single-call summarization instead).

Practical checklist
- OCR and extract page-text + metadata
- Split into chunks, add overlap, assign IDs
- Run chunk summarizer prompt for every chunk
- Index summaries (simple DB or vector index) by topic
- Run topical syntheses, then global synthesis
- Extract canonical refs and run verification pass
- Human review of quotes and final copy

If you want, I can draft the exact JSON schema for chunk outputs or a small script to automate chunking and batching of Claude calls.

Compare Claude and ChatGPT