With the growing acceptance of technology-assisted education, it becomes crucial to create mechanisms to enable students to develop critical skills such as academic writing. English writing presents obstacles to many learners, particularly in relation to students who are short on reliable references while composing accurate and coherent essays. The large language models (LLMs) such as GPT have shown promise in assisting students with writing. But these methods are not without major drawbacks. One of them is notoriously the fact that generative models have been known to generate incorrect or incomprehensible information— referred to as hallucinations—adding falsehoods to the text that might distract students and disrupt their learning experience. In addition, LLMs often lack text coherence in longer inputs, making it impossible to use them in the context of reading comprehension.

This paper presents a learning system that consists of Chain of Thought (CoT) and Retrieval-Augmented Generation (RAG) functionalities integrated into a GPT-based system to enhance such tasks performance. Firstly, the given input task is broken down into small manageably-sized subtasks, which will be solved from one side to the other. Each segment includes making sure external knowledge concerning just that part of the essay has been retrieved, and gives a basis for writing something. The outputs of the individual segments shall be cleaned and assembled into one logical whole, or final essay, at the end. CoT facilitates and guides a model to do the step-by-step reasoning, aiding the students in internalizing and organizing their thoughts and framing logical essays. RAG retrieves accurate external knowledge that provides reliable references to students, minimizing their exposure to factually incorrect or misleading information. So these two reduce the cognitive load of the students by giving them clear, accurate, and coherent information to follow and enabling them to work on their writing skills. By improving the quality of suggestions and reducing errors in the content provided, the system helps students learn more effectively and with greater confidence.

To evaluate the system, we employed two methods:
(1) a comparative evaluation using TrueSkill, where participants compare essays generated by the standard GPT model and the enhanced CoT+RAG GPT model based on accuracy, clarity, and depth,
(2) a pre-and post-test design, where students write 100-word essays before and after using the model, with scores assessed by Grammarly.

The comparative evaluation highlights the superiority of the CoT+RAG-enhanced model, while the Grammarly analysis reveals measurable improvements in students’ grammar, clarity, and vocabulary.

This research contributes to developing AI-assisted learning tools by combining advanced LLM techniques with practical applications. It demonstrates the potential of integrating CoT and RAG to improve not only the quality of AI-generated content but also the writing skills and confidence of English language learners.

Keywords: Retrieval-Augmented Generation (RAG), Chain of Thought (CoT), Factual Accuracy, Large Language Models (LLMs), Writing Skill Development.