1. Prompting for a film, not a shot
'A boy discovers a dragon and they become friends and fly over mountains at sunset' is a story, not a shot. AI video models generate 5-10 second beats. Prompt one beat, not a plot.
2. Skipping the hero frame
Text-to-video has a lower keeper rate than image-to-video across every current model. If your first-shot success rate feels random, you're skipping the frame-lock step.
3. Generating final quality before the concept is locked
Ten Sora 2 flagship rolls to find the composition costs $6. Ten Veo Fast rolls to find the same composition costs $2. Then generate the final once.
4. Wrong model for the shot
Using Sora 2 for dance, Kling for controlled ecommerce orbit, Flux for text-heavy posters. Each model has a sweet spot; using the wrong one triples cost with worse output.
5. Ignoring aspect ratio in the prompt
Default 16:9 for shots that are shipping to a 9:16 TikTok wastes credits. Ask for the ratio you're delivering in.
6. Vague audio
'Cinematic sound' produces generic room tone. Quote dialogue, name foley events, describe ambient specifically. Sora 2's audio quality is directly proportional to prompt specificity.
7. No colour grade
Multiple AI shots in a sequence look mismatched because each was sampled independently. A single LUT applied across the sequence in a NLE unifies them and is the single cheapest polish upgrade in an AI workflow.
Apply the fixes
Put this into practice in the studio — under a minute to your first result.
Apply the fixes →