Models vs products

A model predicts the next token. A product is everything wrapped around that guess: retrieval, tools, logs, a human who can say no, and a story you are willing to defend.

When a demo fails, people blame “the AI.” Name the layer. Was it the weights, the prompt, the retrieved document, the tool schema, or the person who shipped without an eval?

Write the distinction down before you prompt. If you cannot say whether you are changing the model or the product, you are not ready to change either.