The system-prompt hierarchy

When you give a model an instruction, you're not the only voice in the room, and you're not automatically the loudest. There's a priority order to what shapes a model's behavior, and once you can see the layers, you can put your instructions where they'll actually hold.

There are roughly five layers, deepest to most immediate. First, the training data, baked in, the hardest thing to override. Second, the safety training on top of that, the guardrails. Third, the system prompt, the custom instructions, which sit at an elevated attention priority above ordinary conversation. Fourth, the conversation history, whose influence decays with distance, the older a message, the weaker its pull. Fifth, your current message, which gets the highest immediate attention.

Why does the same instruction sometimes stick and sometimes slip?

Because where you put it decides how much weight it carries and how long it lasts. An instruction dropped into a single message lives at layer five, loud right now, but it starts decaying the moment the conversation moves on. The same instruction placed in the system prompt or custom instructions lives at layer three, a persistent frame that keeps applying across the whole session. If a behavior keeps slipping after a few exchanges, that's usually the tell: you put a persistent rule in a temporary place.

The leverage this hands an operator

Use both layers on purpose. Custom instructions set the standing behavioral frame, the things that should always be true no matter what you're working on. Each message gives the focused task for right now. The mistake is trying to run everything through message-level instructions and then wondering why the model drifts back to its defaults twenty messages later. The frame belongs in the frame. The task belongs in the task.

The takeaway: instructions have a priority order, and persistence lives in the system layer, not the message layer. Put standing rules where they stand, and use each message for the job in front of you.