Abstract: Traditional recommender systems either present static ranked lists or engage users in multi turn dialogues. Recent advances in large language models (LLMs) enable a new “one shot” conversational paradigm that bridges the speed of a single search query with the personalisation of dialogue based recommendation. This talk introduces the Guided Choice Framework, an architecture that integrates an LLM driven dialogue front end, a conventional recommender engine, and an uncertainty estimation module that decides whether a clarifying question is required – at most one per interaction. A prototype built on the MovieLens dataset demonstrates how the system parses natural language intent, optionally asks a targeted clarification, retrieves candidate items, ranks them, and generates a concise, explainable recommendation. Six design principles distilled from literature and empirical evaluation that maximise information from minimal input, judicious clarification, contextual personalisation, transparent explanation, grounding in factual data, and calibrated user trust, guide the development of robust one shot CRS. The presentation will discuss performance insights, practical challenges such as hallucination mitigation and handling out of scope requests, as well as future research directions toward adaptive multi turn policies, domain wide applicability, and hybrid LLM recommender models.