November 1, 2026 · NascoreAI team
When an AI video generation comes back wrong, the instinct is to blame the model and re-roll. In production, most failures trace back to the prompt: it described a mood ("cinematic, epic, emotional") instead of a shot. Models are literal. Give them a shot list in sentence form and the keeper rate climbs immediately.
This is the prompt structure we use on our own films. It has five blocks, always in the same order.
Open with the character, described the same way every time: age, build, face, hair. This is your character bible pasted verbatim. If you also attach a reference image, the text still matters — it tells the model which parts of the reference are the identity and which are just that day's lighting.
Spell out the entire outfit in every single prompt, including pieces that are temporarily hidden. This is the most expensive lesson we have learned in production: if a character's apron, jacket or shoulder strap is covered by another object in the opening frame and the prompt does not name it, the model quietly deletes it — and it never comes back in later frames. "Beige t-shirt under a brown apron printed with a logo, red-and-white checkered towel over his left shoulder" survives occlusion. "A food vendor" does not.
One clip, one action. "She turns, draws her sword and charges" is three shots pretending to be one; the model will smear them together. Pick the beat the clip exists for — "she turns slowly toward the sound, gripping the sword hilt" — and let the next clip carry the next beat. Ten seconds of video holds roughly one deliberate action plus a reaction.
Name the shot the way a crew would: "medium close-up, camera slowly pushing in", "wide shot, static camera", "over-the-shoulder, slight handheld sway". Shot size also protects identity: the tighter the framing, the more pixels the face gets, and the better it matches your reference. Save wide shots for moments where the face matters least.
Close with the environment and the light source, concretely: "night market alley, warm light from a single hanging bulb, wet ground reflections". Two or three physical details beat ten adjectives. Light is also a continuity tool — the same stated light source across chained clips keeps a sequence feeling like one moment.
Weak: "A brave young warrior in an epic cinematic battle, dramatic, 8k, masterpiece."
Shootable: "A 24-year-old woman with a scar through her left eyebrow, black hair in a high ponytail, wearing a dark red leather jacket over a grey shirt. She draws a curved sword and steps back into a guard stance. Medium shot, camera tracking slowly to the left. Moonlit ruined courtyard, cold blue light, dust drifting through the air."
Every block is doing a job: identity, costume, one action, camera, light. Nothing is decoration.
Test new prompt structures in Draft mode before spending Final renders on them — on NascoreAI pricing that is $0.59 experiments against $1.15 keepers. When a structure holds the character for three chained clips in Draft, it has earned the Final pass.