Introduction: Clinical text de-identification is essential for secondary healthcare data use but remains challenging due to heterogeneous documentation. Methods: We evaluated LLM-based de-identification of synthetic German discharge letters (GraSCCo) using four prompting strategies with GPT-4o and an additional GPT-OSS comparison. Performance was assessed via precision, recall, F1-score, false positives, and relative text reduction. Results: The baseline setup was not suitable due to excessive text loss. F1-scores reached 0.81 (session-isolated), 0.79 (structured-input), 0.93 (optimized prompt, GPT-4o), and 0.90 (GPT-OSS). Discussion: Results indicate that prompt refinement has a stronger impact on de-identification quality than structural preprocessing. LLM-based de-identification is feasible but requires careful prompt design and validation on real-world data.