Because every prediction is conditioned on the tokens that came before, the prompt is the first thing that shapes the distribution. A vague prompt leaves many continuations plausible, so the answer can wander; a specific prompt rules out most of them early, so the remaining path is narrower and the answer's direction is set from the first token onward. The prompt does not store an answer — it tilts the probabilities that the loop then follows.
How ChatGPT Writes One Token at a Time