This paper presents the technical and creative process of building a dataset of a personal live coding style using MIRLCaProxy, a custom SuperCollider class built on top of the MIRLCa extension. MIRLCa enables real-time sampling of sounds from Freesound with the assistance of machine learning using FluCoMa. We designed an environment that captures eight methods for retrieving sounds in the MIRLCa language, recorded through a live coding journaling approach. This approach aims to predict the next line (next method) from the audio state of the system. Throughout the dataset creation, the required number of actions led to unexpected creative discoveries, transforming the process into a space for sonic exploration. This paper reflects on how the training of the machine learning process becomes a rehearsal space that supports the development of a personal style through constraints. It also explores the role of biases in this context.