Ask Gemini to act as a skeptical analyst that separates verified facts, inferences, and unknowns, keeps dates beside time-sensitive claims, and ends with a decision table you can act on.
O prompt
Act as a skeptical research analyst. Investigate [TOPIC] for [AUDIENCE]. Separate verified facts, reasonable inferences, and unknowns. Use primary sources where possible, include dates beside time-sensitive claims, and finish with a decision table covering evidence, trade-offs, confidence, and next action. If a claim cannot be sourced, say so rather than guessing.
Substitua as variáveis
[TOPIC]The question or decision, stated as narrowly as you can.
[AUDIENCE]Who reads the result — changes depth and tone.
Que modelo utilizar
- Gemini 3.1 Pro — Longer, source-grounded reasoning benefits from the Pro tier; Flash is fine for quick scans.
Dados de entrada de exemplo
- [TOPIC] = "Should a 5-person startup self-host analytics or use a SaaS in 2026?" [AUDIENCE] = "a non-technical founder".
Resultado de exemplo
Só publicamos resultados de exemplo provenientes de execuções reais dos modelos. Este modelo ainda não foi formalmente testado, pelo que não é apresentado nenhum resultado. Execute-o você mesmo com os dados de entrada acima.
Por que esta estrutura funciona
- Forcing a facts / inferences / unknowns split stops the model from smoothing over uncertainty.
- Requiring dates beside time-sensitive claims surfaces stale information.
- A decision table converts prose into something you can actually choose from.
Falhas comuns e correções
The model invents sources or citations.
Correção: Add: "List only sources you can name; mark anything unsourced as UNVERIFIED." Then spot-check the named sources.
Answer is generic.
Correção: Narrow [TOPIC] and add constraints (budget, region, timeframe).
Limitações conhecidas
- Model outputs can still contain confident errors — verify sources before acting.
- Freshness depends on the model and whether web/grounding tools are enabled.
Perguntas frequentes
Does this guarantee accurate citations?
No. It reduces vague hand-waving, but you must still open and verify each named source. Treat unsourced claims as unverified.