1. Two layers, one conversation

By now you can write a strong prompt. This course is about the layer underneath it: instructions a platform includes automatically, in every conversation, without you retyping them. You write them once; they are silently present every time, shaping how each message is read.

That gives every exchange two layers. The persistent layer answers 'who am I dealing with and what are the standing rules?' The per-message prompt answers 'what do you want right now?' Most prompt setups fail by mixing them, standing rules retyped per message, one-off details fossilized into the standing layer.

2. What belongs in each layer

The sorting test is stability. If an instruction would still be true next month on a different task, who you are, how you like answers structured, your definitions of done, your never-do rules. It belongs in the persistent layer. If it is true only for this task, this document, this deadline, this audience. It belongs in the message.

Misfiling costs you in both directions. A stable preference left in messages gets forgotten the one day you skip it. A one-off detail promoted to the persistent layer haunts every future conversation. The model keeps honoring a March deadline in July, because you told it to, permanently.

3. Same mechanism, different wrappers

Platforms expose the persistent layer under different names: an instruction box in settings, per-project or per-workspace instructions, rules files a tool loads from a folder. The wrapper varies; the mechanism is the same, text included before your message that shapes how everything after it is read.

Two practical consequences. Scope: settings-level instructions follow you everywhere, project-level ones apply only inside that project, put each instruction at the narrowest scope that is still stable. Portability: because the mechanism is just included text, a well-written profile moves between platforms almost unchanged. Write for the mechanism, not the wrapper.

4. Every persistent line is paid on every prompt

Persistent instructions change the economics of prompting: a line written once improves hundreds of future prompts, and a bad line degrades hundreds. 'Start with the answer, then the reasoning' pays out daily. 'Always be creative' quietly distorts every task where you needed precision.

So the persistent layer must be short and load-bearing. Every line should pass one question: would you want this applied to every conversation, including ones you cannot foresee? Audit it occasionally. Delete lines you keep overriding and lines too vague to steer anything. Ten sharp lines beat two pages of throat-clearing.