Use these six fixes when the video model drops or ignores details you asked for in your video prompt.
Move the ignored detail earlier. Video models weight the start of a prompt most heavily; details buried at the end get dropped first.
Cut competing descriptors. If you asked for ten things, the video model triaged. Trim the video prompt to the 3-5 details that matter most.
Shorten the overall video prompt. Past roughly 120 words, later instructions are increasingly ignored.
One instruction per element. Conflicting cues ("dark moody night" plus "bright cheerful colors") force the video model to ignore one of them.
Repetition as emphasis: if one detail is critical, reference it twice in different phrasing (e.g., "wearing a red scarf... the red scarf catches the wind").
If nothing works, the video model may not support the concept. Some elements (specific text on signs, exact logos, precise object counts) are known weak points across AI video models — rendering readable text in particular usually fails and is better added in a video editor afterward.