Build a result check into the order. Return ONLY JSON: {"goal":"what you need to get","verify_points":"what facts must be double-checked and how","require_sources":"what to require as evidence (link, quote, screenshot)","freshness":"how to make sure that the data is fresh and not from an old saved copy (cache)"}
The agent is not an all-knowing oracle, but a performer who sometimes makes mistakes in three ways: taking outdated data (an old cache - a saved old copy of the page), confusing similar options (the wrong card), or “making up” a missing fact, because the model tends to fill in the gaps with a plausible one. Hallucination mechanics (when the model produces a plausible but false fact): if something is not on the page, the model can substitute a typical value instead of an honest “not found”. Hence the rule: do not trust numbers without a source. Technique: demand proof - “for each price, provide a direct link and an exact quote from the page.” An agent can hardly falsify a verifiable fact, but a fictitious one is immediately exposed. The second technique is to control freshness: “indicate the date/time the data was received” and “open the current page, not a saved copy.” Insider: ask to show critical facts (price, availability, deadline) literally - with a quote, not a paraphrase. Retelling is the place where mistakes are born. A common rookie mistake is to mistake the agent's confident tone for the truth. Model confidence does not equal accuracy; trust only what is supported by a reference.
Unlock access to submit solutions for instant AI review.