Calibrating large-scale hydrological models is challenging because heterogeneous land-surface processes and parameter interactions complicate error diagnosis and physically meaningful parameter adjustment. Conventional optimization can identify high-performing parameter sets but provides limited process-level explanations. Although large language models (LLMs) offer opportunities for knowledge-guided calibration, explicit hydrological reasoning linked to testable adjustment hypotheses and simulation-based verification remains underdeveloped. We develop a Skill-guided LLM agent for process-aware calibration. The framework organizes hydrological knowledge into reusable procedural guidance through four domain-specific Skills: Diagnosis, Adjustment, Evaluation, and Verification. In a closed-loop workflow, the agent interprets streamflow error signatures, formulates process-level hypotheses, proposes physically constrained parameter adjustments, and evaluates their effects through deterministic simulations. The framework was implementedusingVICv5.1.0 coupled with RVIC in the Jialing River Basin. Under the same model-evaluation budget, it was compared with an LLM-based approach without Skill guidance and the SCE-UA algorithm. Relative to the baseline parameter configuration, the Skill-guided approach increased the Nash–Sutcliffe efficiency (NSE) from 0.690 to 0.716 and reduced mean absolute peak-timing error across the same stations and flood events from 23.22 to 9.89 days. Improvements were not consistent across all hydrological signatures. Calibration trajectories showed how process diagnoses informed parameter interventions and how simulation feedback supported their verification. Local sensitivity analysis showed that parameter sensitivities differed between the initial and optimized configurations. These results demonstrate the feasibility of integrating structured hydrological knowledge, LLM reasoning, and deterministic simulations into a traceable, hypothesis-driven calibration framework for process-based hydrological models.