Someone asks Claude to "review our policy and flag anything risky", without attaching the policy, and gets back a generic list of things policies often get wrong. They conclude Claude is unreliable. Where does the fault actually lie?

The principle — a disappointing answer is worth diagnosing before it is judged, and the first question is always whether the request contained what an answer would need. Two things are commonly missing: the material itself, and the criterion by which the output should be judged.

Why the key is correct — both are absent here. The policy was never attached, so there was no document to review; Anthropic's guidance on uploads exists precisely so that material can be supplied, and a chat accepts documents in formats including PDF and DOCX. And "risky" was left undefined — risky to whom, against which regulation, in which jurisdiction? Anthropic's guidance to new users is that the more specific a request, the better Claude can assist, with the recommendation to refine and ask follow-ups. Given a request with no document and no criterion, a generic list of common policy weaknesses is close to the best available answer: it is what the question, as posed, actually supports.

Why the others are wrong — a more capable model cannot read a file it was not sent, so capability is not the missing ingredient and switching would reproduce the result. Calling it a bug assumes a review of an unsupplied document was possible. And concluding that the whole class of task needs specialist tools abandons a documented strength on the evidence of an attempt that never gave it anything to work with.

Remember this — before judging the answer, check whether the question contained the document and the standard.

Sources — Anthropic's getting started guide and its file upload documentation.