The single most important prompt design principle for journalism is the source requirement. AI tools trained on uncited web content will happily generate authoritative-sounding claims without any source — unless the prompt explicitly requires them. The right phrasing transforms an AI tool from a hallucination machine into a source-retrieval system.
Effective Citation-Forcing Phrases
Add one of these to every factual query: "Cite your sources for every factual claim using this format: [Author/Organisation, Year, URL or Title]." "For every statistic you include, provide the original source document and year." "Do not include any claim you cannot attribute to a specific named source." "If you cannot find a reliable source for a claim, say 'unverified' rather than stating it as fact." The last instruction is particularly powerful — it invites the model to acknowledge uncertainty rather than confabulate.
The Negative Constraint Pattern
Negative constraints are often more powerful than positive requirements. "Do not use phrases like 'studies show' or 'research indicates' without citing the specific study." "Do not round statistics — use the exact figures from the source." "Do not infer causation from correlation without an explicit causal mechanism from a cited source." These constraints directly address the most common AI citation failures.
Verifying AI-Generated Citations
Requiring citations is only half the solution — the citations must also be verified. AI tools sometimes generate plausible-looking citations (author, journal, year) that don't exist. Spot-check citations for any claims that will be published: search the cited source directly and verify it contains the stated claim.