Inpatient medical records which contain clinical narrative information generated from medical procedures in hospital have rich content and unlimited expression capabilities, the information and knowledge implied by which are very useful and important for the proceeding treatment and secondary use of data such as health text mining. In this paper, we proposed a novel method to assistant health practitioners to write narrative clinical text in a more efficient and safe manner. The core technologies beneath this work are named entity recognition (NER) and similarity computation. At sentence level, we used a conditional random field (CRF) method to train a NER model, when doctors type in an entity, several input candidates will pop up for selection; at paragraph level, we used a Gibbs-LDA++ tool and named entities to characterize the topics and key entities of existing records, when doctors create a new clinical text, the patient’s structured data will be used as input to match similar paragraphs, as doctors keep typing in, the matching paragraphs also might change dynamically according to the input content.