A prompt is not a search query and it is not a wish. It is a set of instructions to a camera that does not exist. The closer your words get to a real shot description, the closer the result gets to usable footage.
Most disappointing AI videos are not the tool's fault — they are the prompt's. Beginners describe a topic; the tool needs a shot. 'A video about saving money' gives the model nothing to hold. 'Coins falling slowly into a glass jar on a wooden table, soft window light' gives it a subject, an action, a surface and a light. This guide breaks a working prompt into its five parts and shows the iteration habit that separates usable clips from random ones.
Nearly every prompt that produces usable footage answers five questions, in roughly this order:
A prompt built this way reads like an old steam train crossing a green valley, morning fog, camera tracking alongside, cinematic — and it behaves like a shot, not a lottery ticket.
The most common prompt mistake is a story squeezed into one generation: 'a man wakes up, drinks tea, opens his laptop and smiles at good news'. The model tries to compress all of it into a few seconds and produces a morphing mess in which nothing reads clearly. Real films are built from shots that each do one thing. Ask for the waking, the tea and the laptop as three separate clips, then let your editor join them. Short, single-action requests also fail less: a hand stays a hand for three seconds of pouring far more reliably than through a whole morning routine.
Without camera words, the model guesses — and its favourite guess is a slow, dreamy drift that fits nothing. Camera language is plain and learnable: static shot for a locked frame, slow push-in for growing importance, pan left to right for revealing a space, aerial view for scale, close-up when detail matters. One honest warning belongs here: close-ups are where faces and hands break first, so use them for objects, not people. Match the camera to the job — an establishing shot wants width and stillness; a product moment wants closeness and control.
Half of prompt craft is asking for what should NOT appear. Since generators garble words, end prompts for clean footage with no text, no watermark, no logo. Since faces drift, ask for scenes without people, people seen from behind, or silhouettes when the story allows it. Since busy frames hide errors badly, prefer one clear subject over a crowd. These are not cheats — every real director does the same thing, choosing angles and sets that avoid problems instead of fixing them later at ten times the cost.
Professionals do not write one perfect prompt; they walk towards the shot. Generate, look, change ONE thing — the light, the camera, the action — and generate again. Changing five things at once teaches you nothing about which change worked. And when a prompt finally delivers, save it word for word in a notebook file. A personal library of proven prompts is worth more than any course: next month's video reuses this month's winners, and your style becomes consistent because your language is. The video linked above shows this loop live — prompt, result, adjustment, result.
See the whole process live on screen — every step, exactly as it is done in the video.
▶ Watch Full VideoGeneration is partly random by design — each run starts from a different random seed, so the model interprets your words afresh. That is why generating several takes of a good prompt is normal practice, and why saving the exact wording of a winner matters more than remembering the idea behind it.
Long enough to name the subject, action, setting, camera and light — usually one to three sentences. Very short prompts leave everything to chance. Very long prompts start to contradict themselves, and the model quietly drops the details it cannot fit. If a detail matters, it should be in the first half of the prompt.
Yes, as a starting point. The shot language — subject, action, camera, light — travels well between tools because all of them learned from similar footage. Expect the mood and quality to differ, and expect to adjust. A prompt library organised by shot type, not by tool, survives every tool change.