Recently, deep learning models have achieved a good performance on automatic melody harmonization. However, these models often took melody note sequence as input directly without any feature extraction and analysis, causing the requirement of a large dataset to keep generalization. Inspired from the music theory of counterpoint writing, we introduce a novel musical feature called melodic skeleton, which summarizes the melody movement with strong harmony-related information. Based on the feature, a pipeline involving a skeleton analysis model is proposed for melody harmonization task. We collected a dataset by inviting musicians to annotate the skeleton tones from melodies and trained the skeleton analysis model. Experiments show a great improvement on six metrics which are commonly used in evaluating melody harmonization task, proving the effectiveness of the feature.