Pull the chapter together. A prompt is not a neutral question; it is the run-up the model continues. So how you phrase it, what examples you lay down, how specific you are, all of it shifts which continuation is most plausible, and therefore what you get back. That is the entire reason prompting is a skill, and it follows directly from the next-token machine, not from any magic.

But getting good at the run-up does not make these tools limitless. Shaping the input beautifully cannot fix a problem that lives in the machine itself. The model will still, at times, continue your run-up into something fluent, confident, and completely false, and no prompt fully cures that. The last chapter is about those real limits, told straight.