Loading slide
Most people, without thinking about it, treat a prompt like a search box: type a few keywords, hit go, let the system find the answer. That habit comes from decades of using actual search boxes, and it quietly misleads you here, because a language model is not built like one.
A search engine takes your keywords and retrieves documents that already contain them. It looks things up. The answer existed before you asked; the engine just fetched it. A language model does nothing of the sort. It takes your text and continues it, writing the words that most plausibly follow. There is no shelf it pulls a finished answer from. The answer is composed on the spot, one token at a time, out of whatever you put in front of it.
That single difference is the key to the whole chapter. Your prompt is not a question dropped into a slot. It is the run-up, the opening stretch of text, and the model's reply is just the most plausible continuation of that exact run-up.
Which is why a lazy prompt gets a lazy answer. Type "tell me about dogs" and the most plausible continuation is a bland, general paragraph, because that is what plausibly follows a bland, general opening. The model is not being unhelpful. It is matching the energy of the run-up you gave it. Lay down a specific, structured opening ("compare three dog breeds for a small flat, as a table, with noise level and exercise needs") and the most plausible continuation is specific and structured too.
You are not searching. You are setting up what comes next. Once that clicks, everything else about prompting is just learning to set it up well.