Global-AI-Group-Inc/content-generation-skills / skills/mira-video-prompting
mira-video-prompting
VerifiedWhich of Mira's video models to pick (the live list is in references/selection.md) and how each one treats a prompt: Seedance 2.5, Runway Gen-4.5, Kling 4.0, Kling 3.0 (and Turbo), Gemini Omni Flash, MiniMax H3, HappyHorse, Wan 3.0. Use it before any generate_video call on Mira, when the user asks "which model", "how long can it be", "does it have sound", "промпт для видео", or pastes a video prompt to improve. NOT for stills (mira-image-prompting) or for the camera and light banks themselves (mira-craft).
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 5.3 KB |
| references/action.md | markdown | 4.2 KB |
| references/models/capricornus-1.md | markdown | 9.9 KB |
| references/models/happyhorse.md | markdown | 5.8 KB |
| references/models/kling-4.md | markdown | 5.2 KB |
| references/models/kling-motion.md | markdown | 2.2 KB |
| references/models/kling-turbo.md | markdown | 2.0 KB |
| references/models/kling.md | markdown | 8.2 KB |
| references/models/minimax.md | markdown | 7.7 KB |
| references/models/mira-cinematic.md | markdown | 5.3 KB |
| references/models/omni.md | markdown | 5.7 KB |
| references/models/runway.md | markdown | 5.6 KB |
| references/models/seedance-2-5.md | markdown | 9.9 KB |
| references/models/veo-lite.md | markdown | 5.1 KB |
| references/models/veo-quality.md | markdown | 5.1 KB |
| references/models/veo.md | markdown | 5.0 KB |
| references/models/wan3-prime.md | markdown | 7.9 KB |
| references/models/wan3.md | markdown | 7.8 KB |
| references/selection.md | markdown | 6.3 KB |
| references/textures.md | markdown | 10.8 KB |
| references/universal-rules.md | markdown | 7.8 KB |
references/models/wan3-prime.md
View on GitHub<!-- generated from the registry by comfy_study/comfy_video/build_skills.py; edit the registry, not this file -->
# Wan 3.0 Prime (`wan3-prime`)
| | |
|---|---|
| seconds | 2-30s, any whole second |
| output | 1080p by default; 480p with quality=ld; 720p with quality=sd |
| reference images | 10 |
| audio | yes |
| prompt budget | 8000 chars |
| extend | no |
| draft | no |
| last frame | yes |
| reference clips | up to 5, 15 s each |
| reference audio | up to 5, bound as @AudioN; needs at least one photo or clip alongside |
## When to pick it
the accelerated Wan 3.0: the same long take (30 seconds in ONE generation, one look and one cast), the same shot-numbered storyboard, native audio and 10 photo refs, rendered several times faster; ~1.4x the price of wan3; 2-30s; 1080p
## How it reads a prompt
MODEL: Wan 3.0 Prime (wan3.0-video-prime - the accelerated tier of Wan 3.0 on QwenCloud / Alibaba DashScope; the same model and the same prompt rules, rendered faster), 30 fps, 2-30 seconds, 1080p, with synchronised native audio. It is the ALL-IN-ONE successor to Wan 2.7 and the only model in the line-up that renders up to THIRTY seconds in a single continuous generation - no stitching, no continuation pass, so a whole story holds one look, one voice and one set of characters end to end. OFFICIAL BASE FORMULA (QwenCloud prompt guide): Entity (description) + Scene (description) + Motion (description) + Aesthetic control + Stylization, then the sound layer. Motion must name its amplitude, speed and effect. Aesthetic control = light source, lighting quality, shot size, camera angle, lens and camera movement. Open the prompt with a short comma-separated aesthetic-control header exactly as the official gold prompts do - 'Low angle, medium shot, dusk, mixed light, cool tones with an orange glow, center composition.' - then the entity-and-scene paragraph, then the motion, then the audio. MULTI-SHOT (official, and mandatory once the clip runs past about 6 seconds): Overall description + Shot number + Timestamp + Shot content. Write it in the platform's documented form - a one-line summary of the whole clip, then 'Shot 1 [0-3 s] ...', 'Shot 2 [3-6 s] ...' - covering the exact clip length in the CLIP SPEC with no gap and no overrun, roughly one shot per 3 seconds. On a 30-second clip that means up to TEN shots; do not compress a long clip into four long ones, and do not let any single camera move run past about ten seconds - end the shot and start a new one instead, which is exactly what the shot boundary is for. Name the transition and the camera position for each shot, and repeat the subject's identifying details in every shot so the character, wardrobe and setting stay consistent across the cuts. For a short clip drop the shot labels and write one unbroken take instead. SOUND SUB-FORMULAS (official): Voice = the exact line + emotion + tone + speed + timbre + accent. Sound effect = source material + action + ambient sound. Background music = score + style. Spoken lines go in DOUBLE QUOTATION MARKS with the delivery attached. Synchronise at least one named effect to a visible action, and close with the ambience and the music mood. On a long clip let the sound develop with the story rather than repeating one bed for thirty seconds. IMAGE-TO-VIDEO: with a start frame attached the official formula narrows to Motion + Camera movement - the image already fixes the entity, scene and style, so describe only what moves and how the camera travels ('fixed camera' if it should not move), plus the audio layer. REFERENCE PHOTOS: when SEVERAL photos are attached they reach Wan natively as numbered reference images. The official identifiers are Image 1, Image 2 in upload order - bind each with a concrete noun, 'the woman in the beige coat from Image 1'. Keep every referenced identity exactly as photographed (same face, shape, colours and labels). References do not fix the composition, so describe the full scene yourself; the multi-shot storyboard form remains available and each shot must repeat the reference bindings it uses. GOLD EXAMPLE (verbatim, QwenCloud official docs): 'Steampunk style, dusk, mixed light (gaslight and neon intertwined), side light, cool tones with an orange-red glow, medium full shot, center composition. The camera pushes in slowly, revealing a dark, narrow railway alley in a Victorian city. The bluestone pavement reflects a damp glimmer, and thick fog slowly flows, winding around rusty copper pipes and steam valves. A group of retro-styled robots marches through, emitting small puffs of steam from their joints. Footsteps, the rhythmic "clatter" of metal feet hitting the ground. In the deep background, a giant gear-shaped building rises. A neon sign flickers in the fog, accompanied by a low-frequency electrical hum.' LENGTH: 120-220 words for a single take, and proportionally more for a long storyboard - a thirty-second clip needs every one of its shots described, not a summary. Write only positive description - never phrase an exclusion as 'no ...' or 'without ...'. Wan 3.0 has NO negative-prompt field at all, so an exclusion written anywhere has nowhere to go except into the picture. Do not request on-screen captions or subtitles, and never write resolution or duration codes into the prompt - they are separate parameters.
## Doctrine
Wan 3.0 Prime: the SAME long-take model as wan3 - up to 30 seconds in one continuous generation, one look and one cast throughout, the same storyboard grammar ('Shot 1 [0-3 s] ...' tiling the exact duration), native audio, refs bound as 'Image N', no negative-prompt field - on the accelerated tier that renders several times faster. Pick it instead of wan3 when the user is waiting for the clip; wan3 when cost matters more than time. ~1.4x wan3.
## Format law the critic applies
Wan 3.0 format law: aesthetic prefix of 8-14 terms first; Shot N [x-y seconds] tiling the duration with no gap, one shot per ~3 seconds and up to ten on a 30-second clip; official camera dictionary with adverb speeds; Image N without brackets; voice = speaker + emotion + tone + speed + accent + quoted line; every constraint positive - this model has no negative_prompt field.
## Video references
Attached clips reach Wan 3.0 as reference videos in upload order, counted separately from the photos: Video 1, Video 2. Bind each one to what it lends - motion, camera, grade, rhythm or sound - 'the character from Image 1 repeats the movement of Video 1', 'keep the pacing and cuts of Video 2' - and say what must not carry over. Clips are 1-15 s each, 15 s in total. When a clip is a 3D blockout or playblast (blockout geometry with mannequin, capsule or 3D-model figures in any colours), treat it as a rough guide: bind it to the camera path, the framing and roughly where the characters stand and move, and say that the figures in the clip, whatever they look like, are only stand-ins whose poses and gestures are NOT copied beat for beat - 'the characters do not repeat the movements of the figures in the clip; they look and act as described, in the chosen style, with natural weight and follow-through'. The style comes from the prompt, not from the clip: live action, anime, cartoon or 3D animation alike. Do not ask for 'the framing at each moment' or 'the position of every character', and do not time the actions to the dummies' movements: that makes the model repeat the blockout frame by frame. The blockout's surfaces and simple shapes are placeholders too: flat colours, faceted or noisy surfaces and flat primitive parts become the real materials and forms the prompt describes - say so, and do not claim the blockout has no textures without checking. Copy the figures' motion only when the user wants that exact choreography. The clip is a staging guide (2026-09-15): handled objects and the characters' business with them (the ball, the drink, the phone), gestures, faces, outfits, crowd, plants, props, materials and textures are absent from it on purpose - the prompt writes all of them with their physics (approach, contact, transfer of force, recovery) at real-time speed.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |