If I get up in the middle of the night to pee, I can find the bathroom without turning on the lights. I've learned the route.
You could make me turn on the lights by moving the bathroom. I'd have to pay attention again, but you'd have made the house harder to use.
Good UI design gives me that same feeling of knowing my way around. I can concentrate on what I'm doing because the controls behave as I expect.
If something looks like a text field, I expect to type into it. When it only opens a menu, I have to stop and learn the exception. A tooltip can explain it, but I'd first ask whether it should look like a menu.
What feels intuitive depends on what you've learned. I know my hallway; a guest needs directions. A familiar interface lets people bring experience from other products instead of starting over.
Reinforcement learning offers a rough analogy. An agent learns a policy for choosing actions through interaction and reward. Change its environment and that policy may become less useful. Moving a familiar control can similarly make someone's past experience work against them. Human learning is richer than RL, but both examples make me ask what the learner already knows.
The same question applies to AI tools. Before adding another system-prompt instruction, I'd look at whether the interface makes the intended action recognizable.
Familiarity has limits. Deleting an account deserves more attention than opening a menu. That interruption should help someone understand the consequence of their choice.
Before adding another tooltip, ask what made the wrong action seem reasonable.