A conversational AI toolkit for citizen deliberation in party primary elections

<ns3:p>Primary elections within political parties are increasingly consequential for democratic outcomes, yet voters –both party members and the general public– frequently lack systematic, comparable and trustworthy information about the candidates competing for nomination. Traditional voter information tools, such as Voting Advice Applications, focus on programmatic alignment in general elections and rarely address the specific demands of intraparty contests, where electability, leadership style, integrity, and organisational capital matter alongside policy positions. The recent availability of large language models with web-search capabilities offers an opportunity to develop accessible, adaptable and free analytical tools that systematise public information about candidates while preserving voter autonomy. We developed PRIMARYCAN, an open-source prompt-based toolkit consisting of three components: (1) a main analytical prompt that generates structured candidate profiles in a default accessible mode and an optional professional mode; (2) an external evaluator prompt for auditing outputs in a different model; and (3) a methodological guide. The toolkit produces structured analyses that are comparable across candidates, traceable to public sources, and adaptable to user expertise. Six analytical blocks cover identity and trajectory, discourse and communication, capabilities and performance, consistency and integrity, political capital and electoral potential and (optionally) advanced behavioural indicators. PRIMARYCAN includes safeguards against common pitfalls in algorithmic political analysis: voting verdicts are prohibited; criteria symmetry is enforced; behavioural patterns replace clinical psychometric labels; hindsight bias safeguards are mandatory in retrospective analyses. The toolkit illustrates that conversational AI, when carefully scaffolded with explicit ethical and epistemic constraints, can serve as a useful intermediary for democratic deliberation in primary elections. We discuss limitations (model variability, hallucination risk, structural limits of self-evaluation), pending empirical validation with real users, and pathways for adaptation to other electoral contexts.</ns3:p>