Custom Instructions

A model only knows what it is told. Left to itself it will guess at how you want things done, and it will guess differently every session. Custom instructions are how you tell the node, once, how to behave, so every client picks it up the moment it connects.

How the node serves them

The node keeps a short instructions file. When an AI signs in, the server hands it that file as standing guidance for the whole session. It is the same idea as a system prompt, except it lives with the node instead of being re-typed into each chat. You can read how it is wired in the custom-instructions doc.

What to put in it

Write it like you are briefing a teammate on their first day. A few examples that earn their place:

• Confirm before doing anything that costs money or changes infrastructure.
• Name resources consistently so the cloud stays readable.
• Stay inside the node's own resource group unless asked to go wider.
• If a key is missing, say so plainly rather than guessing.

Updating them

Change the file, redeploy, reconnect. The next session behaves the new way. Because it is one file in one place, the behavior is consistent across ChatGPT, Claude, Grok, and anything else you connect, and there is exactly one thing to edit when you want to change it. (Worth knowing: the instructions sit in the context window on every turn, so keep them tight.)

Instructions versus memory

Custom instructions are static guidance that does not change during use. Memory is different: things the node should remember across sessions, like decisions or preferences. That is a small tool you can add later; instructions are the simple, durable layer to start with.