Treatment effect estimation from observational data faces critical challenges when covariates are partially observed due to resource constraints or privacy concerns. This study introduces a novel framework leveraging world models (e.g., DeepSeek) to address partial observability in treatment effect estimation using a prompting strategy with few-shot in context learning. Specifically, the world model iteratively prioritizes covariate acquisition based on simulated information gain. It dynamically interacts with historical data and domain knowledge to optimize covariate selection under budget limitations, ensuring efficient data collection for unbiased effect estimation. Experiments on a well-known public available dataset (Twins) show the effectiveness of the proposed framework.