Créditos grátis ao criar conta — faça o seu primeiro vídeo de tendência hoje.Começar grátis
Voltar ao blog

The AI video prompt formula: six parts every clip needs

13 de novembro de 2026dimly.art Teamdimly.art Team

A copy-paste structure for AI video prompts — subject, first frame, motion, camera, light, negatives — with a worked before/after example and model notes.

Most AI video prompts fail the same way: they describe a scene ("a dog dancing in a neon studio") and stop. The model fills the rest with guesses — a camera that drifts, light that changes mid-clip, a second dog nobody asked for. The clip isn't broken; it's un-directed.

Here is the structure we use for every prompt on dimly.art's trend pages, and why each part earns its place.

The six parts

  1. First frame. Say where the motion starts: "Use my uploaded photo as the opening frame. Keep the same person, pose, clothes and background." Without this, the model invents its own opening and your face may never appear.
  2. Subject. Who or what the clip is about, stated once, precisely: "a corgi", "the same woman", "a toy robot". Resist adding extra subjects — every extra noun is a chance for a background extra.
  3. Motion. The one thing that happens, in order, with timing: "Over eight seconds the dog dances: two spins, one jump, then holds the final pose." Motion is the product — give it more words than anything else.
  4. Camera. "Static camera" is a legitimate and underrated choice. "Slow push-in", "orbit left", "handheld follow" all work — pick one and name it. Unspecified camera is where most "something feels off" clips go wrong.
  5. Light. "Soft even light", "golden hour backlight", "neon club lighting". Light fixes identity consistency as much as style: changing light makes the model re-imagine the face it already drew.
  6. Negatives. What must not appear: "no text, no logos, no watermark, no other people". Short, concrete, at the end.

You don't need all six in every prompt — a five-second effect can live with three. But when a render comes out wrong, the missing part is almost always one of the six above.

Before and after

Weak:

A grandma dancing in a fancy cafe.

Structured:

Use my uploaded photo as the opening frame. Keep the same person and the same clothes throughout. An elderly woman walks into a luxurious cafe with chandeliers and gold trim, looks around in amazement, then does one delighted little spin and laughs at the camera. Slow push-in from the doorway to a medium shot. Warm evening light, soft and even, photorealistic. No text, no logos, no other people looking at the camera.

Same idea. The second one renders as a directed clip because nothing was left for the model to invent: where it starts, who it is, what happens, how the camera moves, how it is lit, what stays out.

Model notes

  • MiniMax H3 renders 5–15 seconds and scales up to 4K. Best for tight, single-motion clips; keep motion to one idea and let the resolution do the work.
  • Seedance 2.5 renders up to 30 seconds and can generate audio. Longer clips need the motion part written as beats in sequence ("first… then… finally…"), and if you want sound, say what should be heard — "rain on windows, distant chatter" — not just "with audio".

Length scales the credit cost either way, so render at the length the idea actually needs — most trend formats live happily in 8 seconds.

Start from a prompt that already works

Every trend page ships its prompt in a copyable card, plus variations: the age progression page has the full aging timeline written out, pet dance has the spin-and-pose structure, earth zoom has the continuous pull-back wording. Copy one, swap the subject, adjust one motion detail — that edit is usually the whole difference between a template clip and yours.

Open the studio, paste, upload a photo, and render — a 5-second clip is about eleven credits, and a failed render costs nothing.