This paper presents a significant extension of the architecture of IntuiSketch, an Intelligent Tutoring System (ITS), conceived to help students learn in the field of anatomy. This ITS is specifically developed to support learning by drawing using pen-based tablets by analyzing semi-structured sketches on the fly and providing real-time feedback to students. The original system, conceived to process structured drawings, was improved to handle semi-structured sketches based on new recognition and supervision engines. The proposed improvement consists in defining a flexible recognition engine by introducing different types of constraints into the Context-Driven Constraint Multiset Grammar (CD-CMG) formalism. The idea is to categorize constraints into primary and secondary constraints: primary constraints ensure essential accuracy by suspending the recognition process until critical errors have been corrected, while secondary constraints allow minor inaccuracies so that students can continue working without interruption. Based on these new categories of constraints we propose to adapt the production of feedback, i.e. depending on the type of unsatisfied constraints, feedback will be produced in real-time and/or deferred to better adapt to individual learning needs. The system is evaluated on a database of student-drawn anatomical sketches, measuring both recognition accuracy and the relevance of immediate and deferred feedback.