<p>Accurate prediction of the North Atlantic Oscillation (NAO) is of significant societal importance, yet it is hindered by a “weather-climate prediction gap” on the intermediate timescales that bridge these two forecasts. This gap arises because predictability on this timescale is neither primarily determined by atmospheric initial conditions nor fully controlled by slowly evolving boundary forcings like the ocean; instead, it critically depends on the precise capture of ocean-atmosphere coupling processes. Currently, the Earth System Models (ESMs) predominantly use uncoupled data assimilation (UDA) methods, which can trigger “initialization shocks” characterized by dynamical imbalances that degrade forecast accuracy. To address this limitation, this study utilizes a nudging-based weakly coupled data assimilation (WCDA) method to improve 15-day medium-range NAO prediction through the generation of more physically coherent initial fields. Experiments are conducted within the CESM-iCTF model, a modification of the Community Earth System Model (CESM) developed by our previous work that incorporates an artificial intelligence-based convective trigger function (AI-CTF) to better simulate critical air-sea interactions. Contrasting WCDA initialization with traditional UDA, our results demonstrate that WCDA reduces the 15-day NAO forecast error by 8% compared to atmosphere-only assimilation (ADA) and by 30% relative to ocean-only assimilation (ODA). Notably, WCDA forecast closely approaches the performance of the ECMWF operational system while consistently surpassing the uncoupled NCEP and UKMO forecasts. This superior performance stems from the fact that WCDA creates a physically coherent and balanced initial state, which mitigates initialization shock from air-sea inconsistencies and effectively suppresses error growth, thereby translating initial-state advantages into sustained forecast skill. We further find a positive correlation between the relative improvement of WCDA over ADA and NAO event duration, attributed to the persistent feedback provided by sea surface temperatures to the atmosphere via sensible and latent heat fluxes. Furthermore, sensitivity experiments reveal a clear temporal hand-off of predictability. During the period of 1–10 days, predictability is governed primarily by the initial atmospheric dynamical and thermal state (wind fields, temperature, and surface pressure), which is essential for suppressing nonlinear error growth in the NAO’s key dipole regions. During 11–15 days, predictability becomes critically dependent on the slowly evolving ocean temperature, which acts to mitigate the amplification of error as it propagates downstream. These findings highlight the essential role of generating physically coherent coupled initial states for weather-climate prediction and offer clear scientific guidance for advancing next-generation forecast systems and optimizing future observational networks.</p>