Use a verification-first prompt: ask for each claim, one concise statement, exact quote (with page/loc), DOI/URL, and a confidence score. Feed papers in chunks; have the model extract claim→source pairs and mark unsupported claims. Add a human check step: spot-check quotes in the PDFs or Google Scholar. Use Claude for long-context cross-paper synthesis and ChatGPT for quick iterative edits. Output an annotated reference list (APA + DOI).
Compare strengths: Compare Claude and ChatGPT