[Purpose/Significance] This study investigates the integration of narrative thinking into large language models(LLMs) through the Story of Thought(SoT) framework, aiming to enhance AI‘s capacity for understanding and generating humanities data. While existing prompting paradigms such as Chain-of-Thought(CoT), Tree-of-Thought(ToT), and Graph-of-Thought(GoT) have demonstrated considerable success in improving LLMs' logical reasoning capabilities in STEM domains, they remain largely confined to linear or graph-based inference structures that prioritize procedural correctness over contextual richness and semantic resonance. Humanities data, by contrast, demand a deeper engagement with ambiguity, metaphor, causality embedded in temporal sequences, and emotionally charged symbolism—elements that conventional prompting methods struggle to capture. The SoT framework addresses this gap by leveraging narrative as a cognitive scaffold, enabling models to simulate human-like storytelling processes that organize information through situational context, character motivation, causal progression, and affective arcs. This shift from a purely task-oriented reasoning paradigm to a narrative-net-based generation mechanism holds transformative potential for how machines interpret and produce humanistic knowledge. By embedding five distinct narrative techniques—progressive disclosure, branching narration, analogy, analogical reasoning, and metaphor—into the prompting workflow, SoT not only enhances textual coherence and logical explicitness but also fosters the kind of empathetic and culturally attuned expression that lies at the heart of humanities scholarship. This research thus contributes to the emerging field of computational humanities by providing a theoretically grounded and empirically validated methodology for bridging the gap between statistical language modeling and meaning-rich humanistic discourse. [Method/Process] The SoT method systematically incorporates five narrative techniques—progressive disclosure, branching narration, analogy, analogical reasoning, and metaphor—into the prompting pipeline. An empirical experiment was conducted using selected chapters(1, 5, and 18) from the original Zhiyanzhai annotated version of The Dream of the Red Chamber. The DeepSeek-V3 model was employed under both zero-shot and SoT conditions. A three-dimensional evaluation framework was constructed, measuring narrative coherence(40%), logical structure clarity(40%), and emotional semantic depth(20%), each assessed via 5-point Likert scale informed by classical narratology and linguistic theories. [Result/Conclusion] The SoT method significantly outperformed zero-shot prompting across all dimensions. In narrative coherence, So T-generated texts exhibited well-structured temporal sequencing and plot closure, resembling human storytelling. For logical clarity, the method effectively utilized connective markers to articulate causal, contrastive, and progressive relations. In emotional semantic depth, SoT enabled richer symbolic expression and contextualized affective resonance, capturing the tragic undertones of the source text. Weighted average scores for SoT ranged from 2.8 to 4.0, compared to 1.8 to 2.4 for zero-shot prompts. These findings demonstrate SoT's superiority in constructing coherent contexts, identifying causal chains, and generating emotionally charged narratives. [Innovation/Value] This study extends SoT beyond its original application in STEM domains to humanities data processing, proposing a broader definition of SoT as a narrative-centered cognitive framework rather than merely a prompting technique. It introduces a multi-dimensional evaluation system tailored for literary analysis, moving beyond accuracy-based metrics. The work provides both theoretical justification and practical evidence for embedding narrative intelligence into LLMs, offering a viable path toward more human-like comprehension and knowledge production in the digital humanities. [Insufficient/Improvement] Despite the promising results, several limitations of this study warrant acknowledgment and point toward directions for future improvement. First, the empirical experiment is confined to a single literary work—the Zhiyanzhai annotated version of The Dream of the Red Chamber—which, while rich in narrative complexity, cannot fully represent the diversity of humanities data across genres, periods, and cultural traditions. The applicability of SoT to other forms of humanistic content, such as historical archives, legal documents, oral histories, scientific narratives, or cross-cultural mythological texts, remains to be systematically tested. Second, the evaluation relies exclusively on manual scoring by human raters using a Likert-scale framework, which, although grounded in established narratological and linguistic theories, introduces potential subjectivity and limits scalability. Future studies should explore automated evaluation metrics—such as coherence scoring models, discourse parsing tools, or sentiment trajectory analysis—to complement human judgment and enhance reproducibility. Third, the current experimental design treats each narrative technique in isolation within separate prompting conditions, without examining their synergistic or cumulative effects when combined. Investigating how different narrative strategies interact—whether they reinforce, complement, or even interfere with one another—could yield more refined prompting protocols. Fourth, the study focuses primarily on text generation quality, leaving underexplored the extent to which SoT-enhanced outputs improve downstream tasks such as literary interpretation, argumentation construction, or pedagogical applications. Finally, the generalizability of SoT across different LLM architectures and languages beyond Chinese requires further validation. Addressing these gaps will be essential for establishing SoT as a robust and scalable methodology for AI-assisted humanities research.