Why rerolling AI video prompts fails — and how to fix them with better writing

A developer tested an AI video generation prompt ten times and found the repeated failures stemmed from poor prompt structure, not random model variance. The original one-line prompt asked for both a barista bumping an elbow and a coffee spill, giving the model two competing actions to resolve in just five seconds. Analysis revealed two core mistakes: cramming multiple actions into a single clip, and placing the most important action last, where models tend to underweight it. Rewriting the prompt to focus on a single action — the coffee splash — stated first and framed with specific lens and lighting details, produced the desired result in all ten subsequent generations. The author distilled these findings into a practical checklist and an open-source tool called ai-film-crew, which reviews prompts across seven filmmaking roles before generation.