How can large language models (LLMs) support critical design explorations for describing, articulating, and responding to bodily phenomena? This study investigates the integration of code-generating LLMs across design explorations in four different self-tracked or bio-sensed domains: learning, chronic illness, posture control, and knee injury. We synthesise our collaborative research through design inquiries into three themes: 1) forced articulation of fuzzy phenomena, 2) difficulty of escaping local optima, and 3) unexpected fragility of generated sketches. These themes, though potentially applicable in other domain contexts, are grounded in our challenges with code-generating LLMs on embodied and embodiment subjects, with a particular focus on working with technology and data relevant to embodied experiences. We propose two implications for design of LLM-based sketching tools anchored on our design inquiry journeys.