Quick, repeatable workflow:

- Chunking: split papers into ~1,500–2,000 token chunks (~800–1,200 words) with ~200-token overlap; label chunks with ID+source.
- Chunk prompt (use per chunk): “Extract: purpose; methods; 3 key findings (bullets); limitations; exact citation (authors, year, DOI/URL). Output JSON: {id, purpose, methods, findings[], limitations[], citation{authors,year,doi,url}}.”
- Synthesis: merge chunk JSONs by theme, deduplicate, produce 200–400 word theme summaries and a consolidated citation CSV.
- Context windows: if model >32k tokens, feed merged chunks; otherwise do iterative merging (group→synthesize→merge).
- Citation QA: run a DOI/URL validation step (CrossRef/Unpaywall API) on extracted citations.

For long-context merging, consider Claude for careful, large-context synthesis.

Compare Claude and ChatGPT