Abstract: To address the structural knowledge barrier in industrial AI development where domain experts lack programming skills while software engineers lack domain expertise, this paper proposes and empirically validates a novel "domain expert + large language model" development paradigm. Taking a combustion optimization AI-assisted decision-making system for a 660 MW ultra-supercritical unit as a case study, a thermal control engineer with no formal programming background completed full-stack development within3-5 dayss using natural language as the sole development interface. The system comprises a physics simulator, comparative training of five ML algorithms, an ICS data pipeline, multi-objective optimization with 151 protection constraints, and AGC performance assessment, totaling over 12,000 lines of code. Based on empirical evidence including model performance (R2=0.9975) and bug-fix trajectories, five key methodological enablers are identified: domain knowledge encapsulation, physics-data hybrid modeling, incremental error-correction iteration, persistent memory management, and safety-boundary enforcement. This case provides a replicable reference for low-cost, expert-led AI system construction in the power industry. A clone-and-retrain migration to an identical Unit #2 within 3-5 days further validated the paradigm"s cross-unit transferability.