A marketing coordinator at a regional chain of garden centres types 'Write a spring promotion email for our customers' into a new chat. The draft is fluent but features barbecues the chain does not sell and a 20 per cent discount nobody approved. The commercial manager has fixed the offer: 15 per cent off bedding plants for loyalty-card holders, valid for two weekends. What should the coordinator do next?
- ARegenerate the same request several times and keep the draft that comes closest to the real offer.
- BRewrite the request with the approved offer, product range, audience and tone, then check the draft against it. Correct
- CSwitch to the most capable model, since a stronger model is less likely to invent product details.
- DAsk Claude to review its own draft for invented details and confirm that every claim is accurate.
Why A is wrong: Regenerating feels quick and sometimes produces a better draft by chance. It is wrong because the request still lacks the offer and product details, so every version is a guess and the coordinator would be choosing the least wrong one rather than fixing the cause.
Why B is correct: The invented barbecues and discount show that Claude filled gaps the prompt left open. Supplying the fixed offer, what the chain sells, who the email is for and the tone removes those gaps, and checking the draft against the approved terms protects the commercial manager's decision.
Why C is wrong: A more capable model is a natural reach when output disappoints. It is wrong because no model can know the chain's approved offer or stock range unless it is told; the problem is missing context, not reasoning power.
Why D is wrong: Asking for a self-check sounds responsible. It is wrong because Claude has no access to the approved offer, so it cannot tell which details are invented, and its confirmation is not independent evidence of accuracy.