Shot Planning for Text-to-Video

Text-to-video models are genuinely good at short, simple, single-action shots and they fall apart predictably on anything more ambitious. Knowing exactly where that line is, and writing to stay inside it, is a durable skill regardless of which model wins the next model war. The alternative, writing an ambitious multi-action prompt and regenerating it a dozen times hoping for a clean result, wastes far more time than planning the shot down to something the model can actually deliver.

Current-generation text-to-video models share a set of real limits, and they’ll still be roughly true when the specific model names change. Generations are short (a handful of seconds per clip). They handle one clear subject performing one clear action, filmed with one clear camera move, far better than anything with more moving parts. Complex hand and finger work, fast interaction between multiple subjects, and physical contact (grasping an object, sitting down, hugging) are common weak points where models still visibly break down mid-shot.

Write the shot like a line from a director’s shot list, not a paragraph of scene description: subject, single action, camera move, pace, style. “A woman walks through a doorway, camera holds static, natural light, handheld feel” is a shot the model can execute cleanly. “A woman walks through a doorway, picks up a coffee cup, waves to someone off-camera, while the camera pushes in and pans left” is three actions and a compound camera move, and it will visibly blend or break partway through.

When a shot idea is too complex for one generation, the fix is not to keep regenerating the same overloaded prompt. The fix is to break it into two or three simpler shots and plan to cut between them, which is exactly the skill Lesson 6 builds on.