Most existing research on automated essay scoring has predominantly focused on developing feature representations to enhance score prediction accuracy. Notably, earlier studies emphasized hand-crafted feature engineering, while more recent advancements have leveraged neural networks to model various structural features. However, both approaches have treated feature extraction and score prediction as separate tasks. This disjunction has resulted in limited interpretability and overlooked the significant relationship between feedback and scores. To address these challenges, we introduce a novel framework, TBERT, which innovatively integrates prompt-based feature extraction, multi-head attention fusion, and curriculum-style loss weighting. This approach effectively aligns regression scoring with feedback generation. TBERT not only enhances scoring accuracy but also delivers interpretable, rubric-aligned, multi-level feedback, thereby advancing the overall evaluation process with improved interpretability and practical relevance. Comprehensive experiments on public datasets validate the effectiveness and interpretability of our proposed method, particularly emphasizing its remarkable adaptability in narrative essays and long-form writings.