Yes, I’ve seen this intermittently with GPT-5.2-instant in assistant-speaks-first flows usually it’s not the prompt itself but a timing/initial response edge case during handoff. A reliable fix is to enforce a deterministic first utterance (even a lightweight static greeting) while still letting the model handle the rest of the conversation dynamically. You can also improve stability by tightening the system prompt to explicitly require an immediate verbal response on first turn and reducing any tool-calling or reasoning delay on kickoff. In production, most teams end up keeping a fixed first message + model flexibility afterward (hybrid approach), because it avoids the silent handoff issue completely. Incas6e you need help setting that up , you can connect with me @Pedro Lourenco