This paper explores the integration of Large Language Models (LLMs) into Intelligent Tutoring Systems (ITSs) and emphasizes the importance of grounding feedback generation in theoretical frameworks and empirical evidence. It highlights the need for careful and thoughtful design of LLM-based feedback, drawing on insights from previous research on feedback generation in ITSs. The paper outlines four key stages for designing effective LLM-based feedback: triggering feedback delivery, input information for LLMs, content requested in the generated feedback, and the modality for delivering feedback to students. It also discusses the evaluation of generated feedback, emphasizing the importance of metrics such as appropriateness, conciseness, and comprehensiveness. The paper suggests that LLMs can provide adaptive, human-like feedback without requiring extensive training data, but it also warns against over-reliance on these tools and the potential for abuse. The paper advocates for a balanced approach to feedback delivery, considering both the timing and the type of feedback provided. It also emphasizes the importance of incorporating learning science principles such as Bloom's Taxonomy and the Knowledge-Learning-Instruction (KLI) framework to guide the design of effective feedback. The paper concludes by calling for further research into the ethical implications of LLMs in education, including issues such as bias and hallucination, and the need for a comprehensive evaluation of LLM-based feedback systems.This paper explores the integration of Large Language Models (LLMs) into Intelligent Tutoring Systems (ITSs) and emphasizes the importance of grounding feedback generation in theoretical frameworks and empirical evidence. It highlights the need for careful and thoughtful design of LLM-based feedback, drawing on insights from previous research on feedback generation in ITSs. The paper outlines four key stages for designing effective LLM-based feedback: triggering feedback delivery, input information for LLMs, content requested in the generated feedback, and the modality for delivering feedback to students. It also discusses the evaluation of generated feedback, emphasizing the importance of metrics such as appropriateness, conciseness, and comprehensiveness. The paper suggests that LLMs can provide adaptive, human-like feedback without requiring extensive training data, but it also warns against over-reliance on these tools and the potential for abuse. The paper advocates for a balanced approach to feedback delivery, considering both the timing and the type of feedback provided. It also emphasizes the importance of incorporating learning science principles such as Bloom's Taxonomy and the Knowledge-Learning-Instruction (KLI) framework to guide the design of effective feedback. The paper concludes by calling for further research into the ethical implications of LLMs in education, including issues such as bias and hallucination, and the need for a comprehensive evaluation of LLM-based feedback systems.