The prompt

The starting text the model reads from and continues from.

A prompt is the text you type into the playground when you press Generate. It can be a sentence, a phrase, a few characters, or even blank — it is simply the initial context the model reads to understand what you want it to say next. The model does not understand language the way you do; instead it sees the prompt as a sequence of token ids (integers), looks at them one after another, and predicts the single most likely next token given that context. Then it appends that prediction to the sequence and repeats.

The prompt's job is to set the model up to succeed. A well-chosen prompt (e.g., "The capital of France is ") gives the model enough context to produce a sensible continuation; a blank prompt leaves the model guessing, and a contradictory one (e.g., "2 + 2 = 5, so ") may produce nonsense because there is no good continuation for that false statement. The prompt is read through the exact same layers — embedding, attention, RMSNorm, feed-forward — as the tokens the model generates, so anything learned during training shapes both how it interprets the prompt and what it predicts next.

In the diagram, the prompt occupies the leftmost position of the sequence (the first few "slots" in the residual rail). Each token in the prompt flows through the full stack once; only after the model has seen all of them does generation begin. Then the generated tokens append themselves to the sequence, and the model reads the growing context to predict the next one. The prompt is your instruction, your starting point, and the model's only guarantee of sensibility.

Related: The token · The generation loop · The output token