Why Does My AI Character Do Nothing in the Shot?
A video model animates what the prompt says is happening during the shot. Write the action as already done ("having just shoved her") and you get the aftermath. Write it in present tense, give the person doing it the active verb, name where the movement ends, and put emotions after the action. Shots that depend on hitting a target, like a ball going into a net, fail across models, so imply those in the edit.
Why it happens
We rewrote one shot four ways before we saw the pattern. The prompt read, roughly: "he stands on the top step, both hands flat, having just shoved her forward; her body pitches away from him down the steps." The take came back with her stumbling, not being thrown. Four things in the wording caused it:
- Past tense put the action before the first frame. "Having just shoved" left only the recovery on screen, and recovering from a push looks like a stumble.
- The person doing it had no verb. He "stands." Every movement in the prompt belonged to her.
- The movement had no end point. "Pitches away down the steps" ends in mid-air. A throw needs an impact: where the body lands and on what.
- Emotion ate the motion. A paragraph about his jaw, his remorse and a line of dialogue competed with the action. The model spent the shot on his face.
The rewrite came out right on the first try.
Goal-based physics is a second limit. Shots that need an object to travel to a target, like a bicycle kick or a header into a goal, failed on both Seedance and Kling for us. Players kicked the wrong way, the ball sat on top of the net or sailed over. The start frame can be perfect; the model still can't keep the ball on course. Full-body tumbles down stairs are similarly unreliable. A landing on hands and knees works.
How to fix it
- Use present tense. "He shoves," not "having shoved."
- Give the doer the active verb. Whoever causes the action gets the verb, with force: "drives both hands into her back."
- Name the landing. Where the movement ends, and what it hits.
- Put the reaction after the action. "Only then does his chin drop."
- Keep one main action per shot. If the fall still reads soft, cut on the shove and make the landing the next shot.
- Don't ask for target physics. Show the ball already in the net, a celebration, or cut from the dribble to the reaction. Linear motion like dribbling and juggling works.
Example
Before:
Marcus stands on the top step, having just shoved Dana. Her body pitches away from him down the steps. His jaw is set, then a flicker of remorse.
After:
Marcus drives both hands hard into Dana's shoulder blades and shoves. She is thrown off the top step, feet leaving the stone, arms wheeling, and comes down hard on the cobblestones below, landing on both palms and one knee. She stays down. Only then does his chin drop.
How ShowMaker handles it
In ShowMaker you break a scene into shots on the Storyboard and Shot Sheet, so a shove and a landing can be two shots. Shoot with AI writes the motion prompt for a shot from a plain description and shows the credit cost before you render, and the free animatic lets you check the whole sequence before paying for video. See the storyboard docs or the AI storyboard generator.
Related questions
Why does my AI video show the aftermath instead of the action?
The prompt likely describes the action as already done, for example "having just pushed." Write it in present tense so it happens inside the shot.
Can AI video models do sports shots like goals?
Linear motion like dribbling works. Shots that need a ball to reach a target, like a header into the net, failed for us on more than one model. Imply the goal in the edit.
Should I describe emotions in a video prompt?
Yes, but after the action. A long emotion description before or during the action competes with it for the shot's motion.
Every fix on this page is a feature, not a workaround.
ShowMaker Studio is built around these problems — reference-locked casting, location look signatures, coverage-based scenes, delivery-timed captions, and per-shot model routing. Start free with 200 credits: narration over AI stills, no card.