Video Prompting

Video Prompting

Create production-ready prompts for AIMC's generate_video tool or for users to paste into external video models. This is a prompt-director skill, not a separate provider or API path.

Workflow

1. Identify the target model family from the user's words or selected model id. If absent, default finished AIMC generation to google-official/veo-3.1-generate-preview; default prompt-only drafting to a generic cinematic prompt and mention the best-fit model family only if useful.
2. Identify input mode: text, image, keyframes, reference, multivideo, or video. For image/video inputs, label each asset role before drafting.
3. Read references/model-formats.md when model-specific constraints, aliases, or format details matter.
4. Draft the prompt around visible action: subject, setting, motion beats, camera, lighting/style, and audio.
5. Keep generation controls outside the prompt unless the model style benefits from time beats. Return or call with separate parameters: model, duration, resolution, aspectRatio, videoMode, inputImages, inputVideo, enableAudio, negativePrompt, seed.

Default Output

For prompt-only requests, output:

Prompt:
<final prompt>

Recommended parameters:
model: <model id or family>
duration: <seconds, if known>
aspectRatio: <16:9 or 9:16, if known>
resolution: <720p/1080p/4k, if known>
videoMode: <image/keyframes/reference/multivideo/video, if relevant>
enableAudio: <true/false, if relevant>

For finished video requests inside AIMC, draft first, then call generate_video with the final prompt and matching parameters.

Prompt Rules

• Prefer one clear camera move and one clear action arc for clips under 8 seconds.
• For 10-15 second clips, use 2-4 beats or short shot labels.
• Treat image-to-video images as anchors; focus the prompt on motion, camera, performance, and audio instead of redescribing the image.
• Map every reference image/video/audio to a job: first frame, last frame, character, product, scene, camera, motion, rhythm, style, or sound.
• Use concrete physical details: fabric movement, reflections, footsteps, splashes, hand motion, object weight, facial performance, lighting changes.
• Put dialogue in exact quotes and keep it short.
• Do not put model ids, API parameter names, resolution, or aspect ratio inside the prompt text. Put those in Recommended parameters or tool args.
• If using a real-person reference with Sora/Veo/Kling-style identity workflows, warn that authorization or likeness approval may be required.

Model Routing

• Seedance / Agnes: best for reference-heavy prompts, first/last frames, multi-image character/product consistency, music rhythm, short drama, e-commerce, and audio-aware scenes. Use explicit reference mapping and optional time beats.
• Veo: best for cinematic realism, native audio, ambience, dialogue, and controlled camera language. Use [cinematography] + [subject] + [action] + [context] + [style/audio].
• Kling: best for dynamic cinematic motion, action, lip-sync/audio variants, and video-to-video editing on supported models. Keep motion readable and action-driven.
• Sora: best for concise cinematography, strong shot direction, dialogue/SFX separation, and image-anchored motion.
• Wan / Hailuo / Runway / Grok / HappyHorse: use compact subject-scene-motion prompts, one coherent shot, and conservative action complexity.

AIMC Tool Notes

• Use inputImages for image anchors and reference images. The first image is usually the first frame unless videoMode says otherwise.
• Use videoMode: "keyframes" for first/last-frame control.
• Use videoMode: "reference" when images guide subject/style without serving as strict keyframes.
• Use videoMode: "multivideo" for Agnes multi-image blending/storyboard-like input.
• Use inputVideo only for models that support video-to-video editing, currently Kling O1 / Kling Omni families through available providers.
• For Agnes, keep numFrames as 8n + 1 if explicitly set.

Quality Check

Before returning or generating, verify:

• The prompt describes motion, not just a still image.
• Asset roles are explicit.
• Prompt complexity fits duration.
• Audio instructions match whether the model supports audio.
• External parameters match the selected model's allowed durations, input modes, image limits, and resolutions.