Large language models (LLMs) offer incredible potential for complex tasks, but their flexibility presents challenges in defining inputs and evaluating outputs. Researchers are exploring these challenges in the context of creative natural language generation tasks, such as generating citation texts for academic papers. This task presents unique difficulties due to its complex input space and multiple possible outputs. Previous approaches to citation generation have lacked consistency in input requirements and have used limited evaluation measures. This new research introduces a framework for systematically exploring citation text generation using LLMs. It features three core components: input manipulation (varying the input data), reference data (a new dataset based on the ACL Anthology), and a comprehensive evaluation kit. The framework systematically varies input components and instructions, including the cited and citing paper abstracts, citation intent, and example sentences. A novel aspect of the research is the introduction of "free-form" citation intents, providing more nuanced guidance to the LLMs. The results of experiments using two LLMs, Llama 2-Chat and GPT 3.5 Turbo, show that both input components and instructions significantly impact the generated text. Free-form intents and example sentences notably improve performance. Interestingly, the relative performance of different input types remained consistent across various instructions, suggesting that smaller-scale experiments may be sufficient to predict the effectiveness of different input configurations. The research also highlights the importance of employing a diverse set of evaluation metrics. Conventional metrics struggle to differentiate between the two LLMs, while NLI-based metrics reveal performance differences more clearly. These findings emphasize the need for a multi-faceted approach to evaluation in creative NLG tasks. Human studies further reinforce the value of free-form intents and example sentences, impacting both human and LLM performance. The work also reveals interesting qualitative insights: LLM generations were often more verbose but less specific than human-written texts, and the wording of instructions significantly influenced the generated text. While focused on citation generation, this research offers a valuable framework for understanding the interplay of inputs, instructions, and outputs in LLMs applied to other creative text generation tasks. The research has limitations, primarily in its focus on English text from the ACL Anthology and potential information leaks from the generated free-form intents. However, it provides valuable insights and a solid foundation for future work in this rapidly evolving field.