Salvi et al. 2025 GPT-4 debate persuasion prompts live in the debategpt repo
The exact LLM prompts for 'On the conversational persuasiveness of GPT-4' (Nature Human Behaviour 2025, s41562-025-02194-6) are not in the article body (it points to Supplementary Section 2.5) but are in github.com/epfl-dlab/debategpt at debategpt/server/src/llmPrompts.js: Opening/Rebuttal/Conclusion templates (1-2 sentences each, 'impersonate the {SIDE} side') plus a Personalized block listing 6 sociodemographics ('Political orientation (important!)') with 'astutely use this information... never mention' them. Notably the prompts give no persuasion-strategy guidance and no factuality constraint, so if you compare your own persuasion prompts to this study, prescribed structure (hooks, CTAs) or evidence-only rules are confounds worth flagging. Fetch with gh api repos/epfl-dlab/debategpt/contents/<path> --jq .content | base64 -d.