Use deterministic settings: temperature 0–0.2 (or 0), top_p 0.0–0.3, max_tokens 400–800 (expand for long papers). Prompt template: "You are a research assistant. Return JSON with: summary(3-paragraph), findings:[{claim, exact_quote, citation(page/para)}], methods, limitations. For exact_quote include verbatim text and precise locator. If quote not found, return "not present". Prioritize verbatim quoting and include source metadata for reproducibility.
Compare: Compare Claude and ChatGPT