The Critically Endangered Red Handfish (Thymichthys politus) has experienced rapid population decline due to anthropogenic pressure and a rapidly degrading marine ecosystem. Previously abundant in shallow coastal waters around Lutruwita (Tasmania), this small cryptic species now faces extinction with a population of approximately one hundred mature individuals remaining in the wild. To safeguard the species from extinction, a red handfish captive breeding program has been established to bolster the wild population through re-release and reintroduction. To enhance reintroduction efforts, we aimed to improve prerelease conditioning protocols for juvenile red handfish in the captive breeding program. This was trialled by assessing and identifying habitat preferences between five distinct hide types, designed based on the observed sheltering behaviour of wild handfish. Juvenile handfish were monitored by video time-lapse photography and captured footage was analysed using a custom-trained YOLOv26 detection model to quantify animal movement and behaviours. The results suggested that juvenile handfish preferred sheltering in patches of vertical structures and caves with irregular perforations that create mottled shade. Identifying preferred habitat features and behavioural patterns informs preconditioning strategies to improve post-release survival. Machine Learning and AI models provide a low-disturbance tool for investigations of handfish breeding, feeding, sheltering and social behaviours.