Recent advances in web agent technologies have enabled automated web browser operations through natural language instructions. While these technologies show promise for end-to-end test automation, significant challenges remain, such as uncertainty in test execution due to LLMs, increased execution time and cost, and decreased accuracy in identifying operation targets as web pages become more complex. To address these challenges, we propose XTestGen, which generates Gherkin-format test cases and JavaScript step definitions from natural language. XTestGen improves reproducibility by producing deterministic scripts, enhances maintainability through modular step reuse and scenario abstraction, and increases element identification accuracy in complex web pages using hierarchical tree exploration. Our evaluation shows that XTestGen enables abstraction and reuse in test generation and achieves higher accuracy in element identification than naive approaches. A demonstration video is available at: https://youtu.be/sQmsNCPGtPo