Constraint Preservation and Multi-Stage Refinement: An Agent Optimization Framework for Literary Writing

Abstract

The bottleneck of Large Language Models in creative writing lies not in generation, but in goal-directed optimization — improving literary quality while adhering to formal constraints. We propose the first agent optimization framework to explicitly separate hard constraints (formally verifiable, e.g., tonal patterns) from soft constraints (requiring aesthetic judgment, e.g., refined diction) in literary writing. Grounded in two complete case studies — classical Chinese regulated verse prosody correction and prose imagery closure generation — the framework contributes: (1) a formalization of mixed-constraint writing optimization; (2) a dual-axis architecture mandating hard-constraint-first ordering; (3) an invariant preservation principle protecting authorial intent; (4) a divergent-convergent search pattern for soft-constraint problems; and (5) a three-tier knowledge architecture. Case A achieved full prosody correction at a 6.7% modification rate; Case B produced four orthogonally varying closure candidates with a three-dimension evaluation rubric. We discuss applicability boundaries, limitations, and generalization pathways.