video-character

name	video-character
description	Generate videos using EasyBits + Runway Gen-4.5. Handles text-to-video, image-to-video with reference photo, and character-preserved video where the same person/pet/mascot appears across multiple clips with the same face. Use for ALL video generation requests.
allowed-tools	mcp__easybits__video_create,mcp__easybits__character_remember,mcp__easybits__character_list,mcp__easybits__character_delete,mcp__easybits__upload_file,mcp__easybits__get_file,mcp__nanoclaw__send_message,Bash(curl:*)

Video Generation (EasyBits + Runway Gen-4.5)

Generate short videos (2–10s). Backed by EasyBits MCP (video_create) which orchestrates prompt enhancement → still generation (with character refs if provided) → Gen-4.5 animation. Supports character preservation so a specific person, pet, or mascot keeps the same face across multiple clips.

Handles three modes:

• Scene only — no specific subject ("un atardecer en la playa"). Pass prompt alone.
• One-shot photo — user attaches a photo, wants it animated once. Pass prompt + referenceImage.
• Recurring character — user wants a saved subject to appear. Pass prompt + character (slug).

Flow

0. Scene-only requests (no subject)

If the user asks for a pure scene with no specific person/pet/mascot ("un atardecer en la playa", "gato volando por el espacio", "una ciudad futurista"), skip straight to step 3 — translate to English and go. No character or referenceImage needed.

1. User mentions a known character by name

Names like "Toby", "mi perro Luna", "Don Beto", "mi hija María".

mcp__easybits__character_list()

• Match found (by name or slug) → use character: "<slug>" in video_create. Do NOT describe the character's physical appearance in the prompt.
• No match and photo attached → go to step 2.
• No match and no photo → ask: "No tengo guardado a Toby todavía. ¿Me mandas una foto para animarlo?"

2. Photo attached

Find the attachment path from [Image: attachments/img-xxx.jpg] in the conversation.

mcp__easybits__upload_file({ path: "/workspace/group/attachments/img-xxx.jpg" })

Returns a public URL.

If the user mentioned a proper name ("este es Toby", "aquí va mi gata Luna"), ask BEFORE generating:

"¿Quieres que recuerde a Toby para próximos videos? Así no tienes que mandar foto cada vez."

• User says yes → character_remember({ name: "Toby", photos: [url] }) → use the returned slug in video_create.
• User says no or it's a photo without a proper name → use referenceImage: <url> directly in video_create.

3. Translate the user's request to English

Runway performs much better with English prompts. Users speak Spanish — you translate internally.

Rules:

• NO physical description if you're using character (breaks identity preservation).
• YES action + setting + lighting/style.
• One sentence is enough — the tool enhances it internally.

Examples:

Usuario (ES)	prompt param (EN)
"Toby corriendo en la playa al atardecer"	"running along the beach at sunset, cinematic golden hour lighting"
"Mi hija bailando en un escenario con luces"	"dancing on a stage with dramatic concert lighting, slow motion"
"Don Beto manejando un convertible rojo"	"driving a red convertible down a coastal highway, aerial shot"
"un video estilo película de ciencia ficción"	"cinematic sci-fi scene, neon lighting, futuristic atmosphere"

4. Announce BEFORE calling the tool

Critical. Without a text message first, the user sees "escribiendo..." for 2 minutes in silence.

mcp__nanoclaw__send_message({ text: "🎬 Va, dame un par de minutos generándolo..." })

Variations:

• "Listo, empezando — te aviso en ~2 min"
• "Generando tu video, aguanta tantito"

Then call video_create.

5. Call video_create

Defaults for WhatsApp:

mcp__easybits__video_create({
  prompt: "<translated to English>",
  character: "<slug>",            // if character known or just saved
  // OR
  referenceImage: "<url>",        // if using one-off photo (mutually exclusive with character)
  ratio: "720:1280",              // vertical, optimal for WhatsApp
  duration: 5,                    // 5 seconds, good balance
  model: "gen4.5"                 // cinematic quality
})

Override defaults only when the user asks:

• model: "gen4_turbo" — if user says "rápido", "prueba", "borrador"
• duration: 2-10 — if user asks for a specific length
• ratio: "1280:720" — if user says "horizontal" or "landscape"
• ratio: "960:960" — if user says "cuadrado"

6. Deliver the video

Returns { videoFileId, ... }. Fetch the URL, download locally, and send as native video.

mcp__easybits__get_file({ id: "<videoFileId>" })

Returns a signed URL. Download to /workspace/group/ so WhatsApp sends it as a playable video (not just a link):

curl -sL "<url>" -o "/workspace/group/video-<slug-or-id>.mp4"

Then:

mcp__nanoclaw__send_message({
  text: "Aquí está 👇",
  video_path: "/workspace/group/video-<slug-or-id>.mp4"
})

Error handling — always reply in Spanish

Error from video_create	Tell the user
Content policy / moderation rejection	"Lo rechazó el filtro de contenido. ¿Probamos con otra idea menos intensa?"
Character not recognizable	"No logro reconocer bien a [nombre] en las fotos. ¿Me mandas otra más clara?"
Missing photo and no character	"Necesito una foto o un personaje guardado. ¿Me mandas una?"
Timeout / network	"Se cortó la generación. ¿La intento de nuevo?"
Rate limit / quota	"Llegamos al límite de generaciones por ahora. Intenta en un rato."

Rules

• Always translate prompt to English internally.
• Always reply to the user in Spanish (captions, confirmations, errors).
• Never call video_create without sending a text message first announcing it.
• Never include physical description in prompt when using character.
• Never assume the user wants to save a character — ask the first time a proper name appears with a new photo.
• Always call character_list before asking for a new photo if the user mentions a proper name — they may have saved that character already.
• Prefer character over referenceImage when the character is saved — identity preservation is stronger with 1-3 saved reference photos than with a single one-shot.

Managing saved characters

• mcp__easybits__character_list() — list all saved characters.
• mcp__easybits__character_delete({ id }) — delete when the user says "olvida a Toby", "borra a [nombre]".

Cross-skill notes

• Quick animated GIF from a single image (cheaper, no face consistency) → use generate-gif --animate from the gif-gen skill.
• Narrated slideshow from multiple images with voiceover → use video-narrate from the video-narrate skill.
• Static image of a saved character in new scene (no motion) → use generate-lora from the image-gen skill if a LoRA is trained, otherwise face-swap.