Most clips generated from a text prompt look fine in a preview window and fall apart the moment you cut them into a real edit. The reason is rarely the model. It is the shot brief: a prompt that describes a mood rather than a shot gives the model too much freedom, and freedom is what produces the drift you notice later.
Write the subject, the setting, the camera behaviour and the light as four separate statements, then join them. "A ceramic cup on a wooden counter, morning window light from the left, slow push in, no cuts" is a brief. "Warm cosy coffee moment" is a wish. The first gives you something you can re-shoot; the second gives you something you have to accept.
Keep one motion per clip. Two competing movements, a pan and a subject turn for example, read as instability at small sizes, which is exactly where most social clips end up being watched.
Short is not a limitation to work around, it is the format. A three to five second insert covers a transition, a product beat or a reaction shot. If a sequence needs eight seconds, plan two clips and cut between them; the cut hides the seams that a single long generation exposes.
When a take is close but not right, change one variable and regenerate. Changing the prompt, the reference image and the motion at the same time teaches you nothing about which one was wrong.
If the tool accepts reference images, use them for anything that has to look consistent: a product, a character, a room. Words are good at intent and bad at identity. A single reference frame will hold the shape of an object far more reliably than three sentences describing it.
Decide the aspect ratios you need up front. A clip framed for a wide hero banner usually loses its subject when cropped to a vertical feed, so generate the vertical version deliberately rather than cropping the wide one. The same applies to text overlays: leave headroom in the frame so captions have somewhere to live.
Tools like Kling 4.0 are useful precisely at the draft stage: browser-based, free to start, text-to-video with reference images, short clips by design. That makes them a cheap way to test five directions before anyone books a shoot. They are not a replacement for a planned sequence, and they will not fix a brief that was vague to begin with.
Treat each generation as a sketch you can throw away. The value is in the number of directions you can rule out in an afternoon, not in any single clip being finished.