Prompt-Driven Requirements Engineering: Large-Language-Model Agents for Continuous Backlog Refinement

Large-language models (LLMs) have demonstrated near-human proficiency in natural-language generation and understanding tasks. Requirements engineering (RE) remains highly manual, error-prone, and discontinuous-particularly the translation of stakeholder conversations into user stories, acceptance criteria, and traceability artifacts. This paper proposes a prompt-driven RE pipeline where autonomous LLM agents (i) ingest multi-party dialogues, (ii) extract structured backlog items, and (iii) update a knowledge graph that maintains bidirectional traceability. We design a hybrid technique that couples in-context few-shot prompting with retrieval-augmented generation (RAG) and a symbolic rules engine for domain constraints. A 1.4-million-token benchmark composed of 217 anonymized requirements workshops in the aerospace and health-tech domains is released. Experimental results show that the proposed pipeline improves end-to-end backlog accuracy by 31 % and reduces human post-editing effort by 42 % relative to current state-of-practice baselines (manual transcription + Jira templates). Automated traceability link recovery F₁ increases from 0.61 to 0.82 while maintaining 97 % stakeholder satisfaction. The study further reports ablation analyses, latency-throughput trade-offs, and GDPR/PHI compliance measures. We conclude that prompt-driven LLM agents can act as continuous backlog copilots, but emphasize explainability and governance challenges that must be addressed before enterprise adoption.