Text tagging and recommendation are crucial tasks in social media applications. Specific and personalised tagging algorithms are in demand alongside the growing diversity of user needs, content creators, and social media companies. However, the rapid development of these algorithms is hindered by the scarcity of relevant data and the fast-changing nature of topics. Large language models (LLMs) offer a flexible solution with zero-shot classification capability, yet their substantial computational demands make them impractical for frequent, instantaneous tagging. While existing distillation methods aim to address this, they often require additional fine-tuning data or auxiliary data generators. Note that there exists a simple, yet easily overlooked fact that together with a well-formed prompt, a well-trained LLM already contains all the necessary information. In this work, we propose a novel approach that directly extracts a lightweight model from an LLM when we are already given a specific, functional prompt. Using functional projection, we project LLM’s capabilities onto a targeted domain. This extraction process relies solely on numerical float-point data, and our experimental results demonstrate that this method provides a timely and accurate solution to the fast-evolving problem of social media tagging.