A compliance prompt supplies four long policy documents and asks which of them govern a described situation. The answers are usually right but occasionally attribute a rule to the wrong document, and every answer must be traceable to its source. Which combination of changes best serves that requirement?

The principle — Attribution is a structural problem before it is an instruction problem. A model can only tell you which document a rule came from if the prompt made the documents distinguishable in the first place.

Why the key is correct — Anthropic's long-context guidance names three moves and all three are doing work here. First, structure document content and metadata with tags: when using multiple documents, wrap each in a document block with its content and source as subtags. That is what makes the four policies separable entities rather than one long stretch of text, and it is the direct fix for misattribution. Second, ground responses in quotes — for long document tasks, ask the model to quote the relevant parts of the documents first before carrying out its task, which Anthropic notes helps it focus on relevant content and ignore the rest. Quoting before answering also produces the traceable artefact the requirement demands, because the quote and its source travel together. Third, place the query after the material: Anthropic states that queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs, which is exactly this shape of prompt.

Why the others are wrong — Asking for sources is correct in isolation, but leading with the question inverts the recommended ordering and gives up the measured gain. Concatenating the documents removes the boundaries that attribution depends on, which is what is already producing the errors. And instructing the model to read carefully adds a subjective non-constraint of the kind Google's checklist tells you to replace with something concrete.

Remember this — For multidocument work: fence each source, quote before answering, ask last.

Sources — Anthropic, Prompting best practices.