Roteiro e prompts · animação com IA

Como fizemos a cena

Uma cena de luta com 15 segundos, 4 cortes de câmera e som, gerada de uma vez só. Aqui está o método em três passos, com os prompts exatos que usamos. Copie, troque os personagens e teste.

15 suma única geração
4cortes de câmera
US$ 2,52custo do vídeo
Claude (Claude Code)Roteiro, direção plano a plano e montagem
Gemini 3.1 Flash ImageFichas dos personagens e imagens da cena
Kling 3.0 ProVídeo e som, via OpenRouter
ffmpegCorte final, rodapé e cartela
Passo 1

Uma ficha de cada personagem

Antes de qualquer vídeo, gere uma folha com o personagem de frente, de perfil e de costas, sobre fundo neutro. É ela que mantém o mesmo rosto e a mesma roupa em todas as cenas: toda imagem nova usa a ficha como referência.

Ficha do Faísca: frente, perfil e costas
Faíscafrente · perfil · costas
Ficha da Nexus: frente, perfil e costas
Nexusfrente · perfil · costas
Prompt da ficha (exemplo: Faísca)
High-end 3D animated feature film still, Pixar / Disney Animation quality: stylized but believable proportions, expressive faces, subsurface skin shading, detailed cloth and hair, cinematic lighting. Sci-fi futuristic tone. No text, no letters, no logos, no watermark, no UI.

Character turnaround model sheet on a plain neutral mid-grey studio background: the SAME character shown three times side by side, full body head to toe, standing in a relaxed neutral pose: front view (left), side profile view (center), back view (right). Identical outfit, colors and proportions in all three views. Even soft studio light plus a subtle glow from the character's power. A tall white man around 30, 1.90 m, broad shoulders, striking bright green eyes, messy brown quiff hair, short stubble beard, confident half smile. Wears blue jeans, a dark brown leather jacket and a plain white t-shirt with no print, brown boots. Power: electricity — small teal-green lightning arcs crackle along the jacket zippers and between his fingers.
Passo 2

Só a primeira e a última imagem

Não desenhe a cena inteira. Gere apenas o quadro de abertura e o quadro final, no formato do vídeo (9:16), usando as fichas como referência. Trave a geografia: heróis sempre à esquerda, vilão sempre à direita. A última imagem de uma cena vira a primeira da próxima, e é isso que emenda um trecho no outro.

Primeira imagem: os três heróis à esquerda encarando o Ōkami à direita
Primeira
Última imagem: close do Ōkami sorrindo com as veias roxas
Última
Prompt da primeira imagem
High-end 3D animated feature film still, Pixar / Disney Animation quality: stylized but believable proportions, expressive faces, subsurface skin shading, detailed cloth and hair, cinematic lighting. Sci-fi futuristic tone. No text, no letters, no logos, no watermark, no UI.

Screen geography is fixed: the three heroes are always on the LEFT side of the frame, Ōkami always on the RIGHT side, on a rain-soaked Shibuya-like crossing at night. One single full-frame vertical image, no borders, no text, no letters, no signs with readable words. Vertical 9:16 wide shot, 24mm, eye level. Left third: Faísca in front, Nexus and Brasa just behind him, all three facing right, powers starting to glow (teal sparks in Faísca's fists, cyan code around Nexus, small orange-magenta flames in Brasa's hands). Right side: the colossal 2.5 m Ōkami, a head and shoulders taller than Faísca and twice as wide, standing calm at the edge of the violet-glowing crater, facing left toward them. Heavy rain, wet asphalt reflecting pink and blue neon, broken giant screens flickering overhead.
Prompt da última imagem
High-end 3D animated feature film still, Pixar / Disney Animation quality: stylized but believable proportions, expressive faces, subsurface skin shading, detailed cloth and hair, cinematic lighting. Sci-fi futuristic tone. No text, no letters, no logos, no watermark, no UI.

Screen geography is fixed: the three heroes are always on the LEFT side of the frame, Ōkami always on the RIGHT side, on a rain-soaked Shibuya-like crossing at night. One single full-frame vertical image, no borders, no text, no letters, no signs with readable words. Vertical 9:16 close-up, 85mm, slightly low angle. Ōkami's face and right shoulder in three-quarter view, turned toward frame-left (toward the unseen heroes). A calm, almost fatherly smile. Violet veins glow and pulse under the skin of his neck and forearm; on his forearm a burn wound has just closed, the last purple light fading over new skin. Raindrops bead on his cracked round glasses, pink and teal neon reflected in the lenses. Metal plate on the shoulder wet with rain.
Passo 3

Roteiro plano a plano para a IA de vídeo

Entregue as duas imagens ao Kling 3.0 Pro (primeiro e último quadro) com um roteiro dividido em planos: tempo, lente, ângulo, ação e consequência de cada um. Ele preenche o meio e faz os cortes sozinho.

TempoO que acontece
0–4 sFaísca dispara um raio no peito do Ōkami, que desliza para trás no asfalto molhado.
4–8 sNexus hackeia os telões; o maior despenca sobre o Ōkami.
8–11 sBrasa, no skate, acerta uma bola de plasma no braço dele.
11–15 sClose: a queimadura fecha com luz roxa e ele sorri, calmo.
Prompt do vídeo (Kling 3.0 Pro)
[Character A: Faísca — tall white man, green eyes, brown quiff, stubble, brown leather jacket, white t-shirt, jeans; teal lightning]
[Character B: Nexus — Black woman, big afro, cyan holographic visor, gold hoops, cropped iridescent jacket, black cargo pants; cyan code]
[Character C: Brasa — skinny 15-year-old boy, curly hair, backwards green cap, green hoodie, shorts, skateboard; orange-magenta plasma]
[Character D: Ōkami — 2.5 m giant Japanese scientist in his 50s, grey hair tied back, cracked round glasses, metal plates on shoulders and forearms, glowing violet veins, torn white lab coat with ID badge; calm, almost gentle]

Master intent: stylized 3D animated feature film, round one of a street fight on a rain-soaked neon Shibuya-like crossing at night. Characters A, B and C always stay on the LEFT of frame, Character D always on the RIGHT; every attack travels left to right. Each hit lands harder, and Character D is still standing. Character D is a head and shoulders taller than Character A and twice as wide.

Shot 1 (0-4s). Medium shot from behind Character A's left shoulder, 35mm. Character A plants his boots in a puddle and fires a thick teal-green lightning bolt from both hands into Character D's chest plate; Character D slides back, heels gouging the wet asphalt, sparks skipping across puddles.
Shot 2 (4-8s). Low wide shot, 24mm, tilting up. Character B throws both arms up; cyan code races up into the giant broken screens, they flash cyan, and the largest screen falls onto Character D, bursting into glass and sparks.
Shot 3 (8-11s). Medium tracking shot, 50mm, moving left to right. Character C kicks off on his skateboard and hurls an orange-magenta plasma ball that explodes on Character D's right forearm, leaving a black smoking burn.
Shot 4 (11-15s). Slow push-in, 85mm, close-up on Character D's face and forearm. The burn closes with soft purple light, violet veins pulse, rain beads on his cracked glasses, and he smiles calmly toward frame-left. Final image: his calm smile.

Hard cuts, each shot a different angle.
Lighting. Pink and blue neon reflected in wet asphalt, strong rim light from each power, violet glow from the crater.
Audio. Heavy rain throughout, lightning crack, glass shattering, plasma whoosh, a low violet hum at the end. No dialogue.
Prompt negativo
subtitles, on-screen text, readable signs, logos, extra characters, monsters, robots, transformation, distorted hands, extra fingers, melted face, music, dialogue

Configuração: 9:16, 720p, 15 segundos, som ligado, primeira imagem como first frame e última como last frame.

Sendo honesto

O que deu errado no caminho

  • Clipes curtos soltos parecem colagem. 22 clipes de 6 s (Veo 3.1 Lite) ficaram bonitos, mas com cara de "monte de cenas juntas". O que resolveu foi gerar a sequência inteira de uma vez, amarrada pela primeira e pela última imagem.
  • O Seedance 2.5 recusou os personagens. O filtro achou que as imagens podiam ser de pessoas reais, mesmo sendo 3D.
  • O Kling aceita no máximo 2.500 caracteres de prompt. Escreva enxuto: um bloco curto por personagem e uma linha por plano.
  • A IA inventa. Vilão virando monstro, figurantes do nada, letreiros com texto falso. Escreva no prompt o que não pode mudar: "nobody transforms, no extra characters, no text".

Personagens e cenário: Gemini 3.1 Flash Image. Vídeo: Kling 3.0 Pro via OpenRouter. Direção e prompts: Claude, com a skill de vídeo de Serge Shima (smixs/visual-skills, CC BY 4.0). Ver o projeto Power Kids →