| （英） |
This paper reports on the process and effectiveness of migrating an existing WordPress site to a Git-based headless CMS that leverages Large Language Models (LLMs). The system architecture we adopted manages content as JSON files within a Git repository instead of a database and builds a static frontend using Next.js. This approach resolves the challenges associated with WordPress, such as performance degradation, high maintenance costs, and security vulnerabilities. Daily content updates are handled conversationally with an AI agent (Claude Code CLI). Notably, we established an efficient editing workflow by having the AI generate and reference its own instruction manual, overcoming the context-persistence challenges of stateless interactions. Evaluation results show a significant improvement in site performance, with the average Lighthouse score increasing from 73 to 94, and a substantial reduction in content creation time. Furthermore, our findings reveal a novel advantage: the AI agent can dynamically generate unique HTML and CSS for each article, enabling expressive designs unattainable with traditional template-based systems. |