The prompt skeleton
Every strong prompt in this library shares a skeleton: CAMERA + SUBJECT + ACTION + LIGHTING + MOOD + DURATION. Write the skeleton, then replace one slot at a time. The generated clip is your diff checker — the model shows you what each choice actually changed.
Subject specificity
Models reward concrete nouns. "A mantis" is weaker than "a praying mantis, compound eyes reflecting an LED strip", because the second supplies texture the model fills instead of inventing. Specificity is the cheapest quality lever in prompting.
Camera placement
Camera words are verbs, not adjectives. "Handheld" changes the whole feel; "static tripod" is a different contract. Keep the camera instruction at the start so it frames everything after it. A clip that drifts often just lacked an explicit camera seat.
Lighting = temperature
Lighting words set the emotional temperature before the action even starts. Warm = intimate, cool = distance, neon = friction, volumetric = awe. If a clip feels emotionless, the prompt usually named no light source at all.
One hero motion
Pick one dominant motion and protect it. A clip with tracking, slow motion and a zoom reads as mush. The corpus’s best clips state one camera move, one subject motion, and let the model animate the rest.
Length contracts
Under a model’s max, length is a promise, not a suggestion. State "8-second clip" explicitly. The library records the author’s posted length because the prompt and the clip are born together — mismatch usually means a crop, not a lie.
Aspect as composition
9:16 for faces and speech, 16:9 for scenery and motion, 2.39:1 for cinema, 1:1 for cards. Models compose to the native ratio. Cropping a 16:9 output to vertical loses half the picture — say the ratio in the prompt instead.
Negative instructions
State what you do not want the way the model hears it: "no text, no watermark, no people". "Without" and "no" behave differently across models — test both against the same clip and keep the one that holds.
Reference transforms
Reference-to-prompt ("in the style of driving at night through rain", "shot like a dashboard camera") is the fastest route to a specific look. The technique page explains how to steal a look without stealing a scene.
Iterate on the clip, not the words
A clip costs one generated video; staring at a prompt costs your time and learns nothing. Run the prompt once, then edit one token, run again. Three runs with the clip visible beat an hour of second-guessing the wording.
Putting it together
A finished prompt is the skeleton warmed by every technique at once: a specific subject, one hero motion, an explicit light source, a stated length and ratio, and a negative pass. Here is what that looks like in practice:
Handheld shot, close up on a street food vendor slicing a mango in a rainy neon market at night, warm halogen glow reflecting off wet tiles, slow motion, steam rising off the grill, no people in the background, volumetric haze, 8-second clip, 9:16.