An internal assistant answers questions about your company's products. It is fluent and confident, but it invents product details that have never existed, and no rewording of the prompt has reduced the rate. The product catalogue is internal and changes weekly. What does this pattern indicate?

The principle — Prompting shapes how a model uses what it has. It cannot supply what the model was never given, and no amount of rewording crosses that line. Recognising the boundary is what stops weeks being spent on the wrong lever.

Why the key is correct — Every fact in the diagnosis points at absent information rather than unclear instructions. The catalogue is internal, so it was not in any training data; it changes weekly, so even a model trained on it would be stale; and rewording has already failed repeatedly, which is the evidence that the constraint is not linguistic. Google's guidance treats this as a tooling question, stating that grounding connects a model to real-time content and should be enabled whenever it may need to know obscure or recent facts, and framing tool use as the way to avoid hallucinations where a model might otherwise produce incorrect responses. OpenAI describes the same move for private data — adding relevant context to the request, the technique commonly called retrieval-augmented generation, in order to give a model access to proprietary data outside the set it was trained on. Anthropic's overview supplies the general permission to stop tuning wording: not every failing result is best solved by prompt engineering.

Why the others are wrong — Further iteration is the correct default for most faults but has been tried and cannot manufacture missing data. Self-checking has nothing external to check against, so the review inherits the same gap as the answer. And instructing the model to hedge improves how a wrong answer is presented without making it right.

Remember this — If rewording repeatedly fails, ask whether the model could possibly know the answer. Sometimes the fix is data, not language.

Sources — Google, Prompt design strategies; OpenAI, Prompt engineering.