Direct sequencing of clinical specimens is now a trending approach for antibiotic resistance detection in Mycobacterium tuberculosis (MTB). Comparing with the golden standard phenotypic drug susceptibility test (pDST) and targeted sequencing with clinical isolates that requires 14-day Mycobacteria Growth Indicator Tube (MGIT) culture, it can greatly reduce the time to report from weeks to a few working days. This is crucial for the choice of correct antibiotics during medical prescription, especially for patients carrying resistant MTB or intolerant of certain side effects of antibiotics. It also helps to understand the antimicrobial resistance (AMR) profile within the community and take action to control the spread. Unlike target sequencing with pure clinical isolates, variant calling from sequencing data of clinical specimens was susceptible to different source of contamination and background nasal/oral flora interference. These potential risks should be highlighted and should be addressed before putting direct sequencing into clinical service. In this study, a new target sequencing workflow was developed, including a set of multiplex PCR primers and a bioinformatics pipeline. The sequencing panel allowed the rapid direct sequencing of clinical specimens. With the synergistic power of MegaPath (a software was designed for detecting low similarity to the reference genome) and Clair-ensemble (one benchmarking deep neural network based variant caller achieves high precision, recall and speed in variant calling for NGS, PacBio, and ONT sequencing data), minimized background nasal/oral flora interference and accurate variant calling were achieved. In short, this study was not only about how a targeted sequencing workflow for antibiotic resistance detection was developed, but it also revealed the advantages and the challenges of direct sequencing of clinical specimens that could be a reference for future direct sequencing development in infectious diseases.