Background: User story quality assessment is a recurrent challenge in Agile Software Development, as deficiencies in clarity, structure, and intent may negatively affect downstream activities such as planning, estimation, testing, and implementation. Prior work has explored both rule-based approaches grounded in the Quality User Story (QUS) framework and learning-based techniques using Large Language Models (LLMs). Objective: This study takes an empirical step towards automated user story quality assessment by comparatively analyzing rule-based and LLM-based approaches for a subset of QUS criteria, focusing on syntactic and pragmatic aspects operationalized by AQUSA. Method: We conduct a comparative evaluation of AQUSA and multiple contemporary LLMs, including GPT-family models and DeepSeek, using 182 user stories from three industrial and open-source projects. Effectiveness is assessed through precision, recall, and F1-score, complemented by a qualitative analysis of feedback clarity and usefulness. Output stability is examined by repeating each LLM-based assessment five times per user story. Results: The results indicate that both approaches are effective in identifying syntactic and pragmatic quality defects, while exhibiting distinct behavioral and variability patterns across models and executions. Conclusion: By consolidating empirical evidence on syntactic and pragmatic QUS criteria, this study provides a methodological baseline for a broader research agenda on automated user story quality assessment, supporting future investigations that extend toward semantic criteria and integrated assessment strategies.