How to ask AI so it cannot hide the gaps

You asked a fair question and got back a beautiful wall of prose. It flows, it sounds authoritative, and there is nothing to grab onto: no source named, no claim separated from the next, nowhere to start checking even if you wanted to.

That wall is not an accident. An open-ended question invites the model to write the most plausible essay it can, and plausible essays hide their gaps in the mortar between sentences. The fix is not a smarter model. It is a question shaped so that every claim arrives as a separate, sourced, checkable brick.

Start by naming what the AI may draw from: "answer only from the attached contract", "only from the 2025 policy manual", "only from the sources you can cite". Scoping does two jobs at once. It shrinks the space the model can improvise in, and it makes the next demand enforceable.

Then require the receipt: "support every claim with a citation to the specific section or passage." Per-claim is the important part. A bibliography at the end lets one real source launder ten unsupported sentences. One citation per claim means every sentence either has evidence attached or is visibly naked, and visibly naked is exactly what you want to see. Peer-reviewed testing found GPT-4 fabricating 18% of the references it generated freely, which is why the citation must point INTO your scoped sources, not out into the model's imagination.

This is the element almost everyone skips, and it changes the whole game. Add one sentence: "if the sources do not contain the answer, say so explicitly instead of answering from general knowledge."

Without that permission, the model faces a forced choice between your question and the truth, and it is built to choose your question. With it, the honest gap becomes an allowed output, and you will be surprised how often it appears. Every "the sources do not address this" you receive is a fabrication that did not happen and a research gap you now know about.

Set the time boundary: "as of this date", "using the version attached, not any later amendment." Stale answers are a quiet failure mode; a cutoff makes time explicit instead of assumed.

Then fix the shape: "answer as a numbered list, one claim per line, citation after each claim, open questions in a separate section at the end." Shape is what makes verification fast. A numbered list with per-line citations can be checked in the order it arrived; the open-questions section forces the model to put its uncertainty somewhere you can see it instead of smoothing it over.

The full template, ready to copy: "Using only [sources], as of [date], answer: [question]. Support every claim with a citation to the specific passage. If the sources do not contain the answer, say so explicitly. Format: numbered claims with citations, then a section for what the sources leave open."

Honesty is now the invitation, but in a general chatbot it is still only an invitation. The model may comply beautifully today and drift tomorrow, and the citations it produces still need the 90-second check, because a request is not an enforcement mechanism.

Tools built around grounding turn these five elements from etiquette into architecture: Tatsulok scopes every answer to your documents and collections by design, attaches the passage link to every claim automatically, and treats "your sources do not contain this" as a first-class answer rather than a favor. The prompt pattern is still worth mastering, because it travels with you everywhere.

This week's drill, no signup needed: take a question you asked an AI recently and re-ask it with all five elements. Compare the two answers side by side. Count the claims you could check in the first, then in the second. The difference is what the pattern buys you.