Objectives The formulation of earthquake emergency response plans is a highly standardized task that demands both procedural rigor and contextual adaptability. Existing methods often struggle to balance consistency with the need for situational specificity and actionable detail. This study aims to address these challenges by proposing an intelligent workflow for the automated generation of earthquake emergency plans. The objectives are threefold: (1) To enhance generation efficiency through automation. (2) To ensure procedural compliance and logical consistency; and (3) to improve the contextual relevance and implement ability of generated plans.

Methods To achieve these goals, we propose a four-stage intelligent emergency plan generation framework that integrates large language models (LLM), intelligent agents, and knowledge graph. The workflow consists of four sequential stages: User intent parsing and structured input construction, agent-driven post-earthquake risk analysis, knowledge-enhanced response strategy generation, and LLM-driven format optimization and quality evaluation, enabling end-to-end automated plan generation. To support this workflow, we design a three-tier emergency plan template system covering provincial, municipal, and county levels, and construct a comprehensive emergency knowledge graph containing 117 915 nodes and 182 586 edges, representing key entities such as risk types, response measures, resource configurations, and organizational responsibilities.

Results Experiments were conducted using 20 city- and county-level earthquake emergency plans as reference cases. A six-dimensional evaluation framework was designed to compare the proposed workflow method against direct LLM generation. Results demonstrate that the workflow method significantly outperforms direct LLM generation in key metrics such as completeness, feasibility, and coherence, producing more structured, logical, and actionable outputs. Further case analysis highlights the workflow's advantages in resource detailing, paragraph structuring, and content consistency, underscoring its effectiveness in real-world scenarios.

Conclusions We present a novel, knowledge-augmented and agent-assisted framework for intelligent emergency plan generation, offering a practical and scalable solution to a traditionally manual process. By combining structured template, semantic knowledge graphs, intelligent agent reasoning, and LLM-based generation and evaluation, the proposed workflow ensures both procedural compliance and real-world applicability. An interactive planning tool has been developed to facilitate real-time use by emergency management practitioners, enabling customizable, region-specific, and ready-to-implement emergency plans. Future work will extend the system's capabilities to other disaster types and integrate predictive analytics for dynamic risk evolution modeling.