A prompt pastes in a long support article and asks a question about it, but the model's answer keeps drifting into general advice that is not in the article. Which addition most directly fixes this?

The principle — Supplying material and bounding the answer to it are two separate instructions. A model given a document alongside a question can legitimately draw on either the document or its general knowledge, and it will often do both.

Why the key is correct — Google's worked example does exactly this. The prompt that fixes a generic answer opens with the instruction to answer the question using the text below and to respond with only the text provided, and the response then comes straight from the supplied material. The scoping sentence is what converts pasted context into a boundary.

Why the others are wrong — A link is inert text without a retrieval tool, and the article is already in the prompt anyway. Personal framing changes register rather than sourcing. And a longer answer typically means more of the general advice that was the problem.

Remember this — Pasting the source says what to read. "Answer using only this" says what not to.

Sources — Google, Prompt design strategies.