Although many current systems do not clearly translate evidence-based intervention principles into explicit software design and interaction logic, conversational agents are being increasingly investigated as interactive tools to support people with autism spectrum disorder (ASD). This paper presents an engineering-oriented framework for designing a conversational system based on Pivotal Response Treatment (PRT), a naturalistic intervention approach that emphasizes motivation, self-initiation, self-regulation, and responsiveness to multiple cues. Instead of emphasizing clinical results, the proposed work embeds PRT principles into a modular, state-based software architecture and formalizes them as functional and non-functional system requirements. To ensure predictability, transparency, and reproducibility, the framework specifies emotion-aware interactions, deterministic session management, adaptive difficulty regulation, explicit interaction modes, and natural reinforcement mechanisms. To facilitate naturalistic dialogue while maintaining complete control over the interaction flow through deterministic logic, an optional, constrained language generation component is included. To demonstrate how therapeutic principles can be operationalized as enforceable engineering constraints, this paper describes the system requirements, architectural design, interaction model, and illustrative usage scenarios. In the context of educational and therapeutic engineering, this study attempts to offer a replicable and expandable reference architecture for the creation of adaptive conversational systems for ASD.