When an AI answer matters, asking the same system whether it is confident does not turn the answer into evidence.
An assistant may correct a mistake, repeat it more confidently, or introduce a new one. Conversation can help inspect reasoning, but verification needs an independent basis.
Classify the output first
Mark important statements as facts, calculations, interpretations, recommendations, or creative choices. Each category needs a different check. Facts need suitable sources. Calculations need reproduced inputs and formulas. Interpretations should be compared with original material. Recommendations must reveal their criteria and assumptions.
Verify consequential claims first
Ask what would cause the greatest harm, cost, embarrassment, or rework if wrong. Check those claims before polishing minor details. For information that changes—prices, laws, product features, schedules, policies, and software behavior—record the source date.
Trace citations to the original
A convincing citation may be fabricated, misattributed, or attached to a claim it does not support. Open the source. Confirm title, author or organization, date, method, and the passage or data supporting the exact claim. If the original is unavailable, keep it labeled as a lead rather than evidence.
Reproduce calculations
Ask for the formula and intermediate steps, then calculate independently with a calculator, spreadsheet, code, or another appropriate method. Check units, rounding, sample size, and input values.
Make uncertainty visible
Use clear labels: supported, contradicted, unresolved, or dependent on an assumption. If sources disagree, compare definitions, populations, methods, dates, and incentives before deciding why.
Claim · type · importance · source · source date · supporting evidence · limitation · result · final revision.
A stronger follow-up
Do not rewrite yet. List the factual claims, calculations, assumptions, and recommendations. For each, state what independent evidence would verify it. Mark anything unsupported as [VERIFY].
This prompt is still analysis, not proof. Its value is that it creates a map for human verification.
For a complete source-first workflow, explore Research & Decisions with AI.