Set up Custom Instructions once so every new chat already gets you

Why this works

Custom Instructions are static text injected as system-level context at the start of every new conversation, which is mechanically different from Memory: Memory accumulates facts automatically over time from what you say across chats, while Custom Instructions are a fixed profile you write once and control directly. Splitting the "about me" facts from the "how to respond" rules mirrors the real two-field structure the settings page ships, and keeping them in separate blocks matters because a model asked to both know a fact and follow a style rule from one undifferentiated paragraph will sometimes recite the fact back as if it were being asked to, rather than quietly using it as background. The instruction to ask one clarifying question rather than guess specifically targets the first message of a brand-new thread, the one point where the model has no conversational context yet to catch a misread request, so it defaults to running with an assumption unless told otherwise. Naming a concrete length target with an explicit "do not pad" clause also closes a known side effect of length instructions alone: told only to keep answers short, a model will sometimes hit that target by omitting substance rather than saying less with the same content.