Ai Video Production

---
name: ai-video-production
description: Route AI video production work to the right skill for ad analysis and dialogue-led recreation, Blender previs, Seedance 2.5, MiniMax H3, Wan 3.0, supporting GPT Image 2.5 still images, or local video metadata cleanup. Use for multi-stage production, reference planning, or choosing among these skills; go directly to the named skill for a simple specific request.
---

# AI Video Production

Identify the requested deliverable and current stage, then use the smallest relevant set of skills from the [skill catalog](references/skill-catalog.md). Complete the authorized stage; a routing decision is not the deliverable.

## Select by task and output medium

1. **Ad reference analysis or dialogue-led recreation:** Use `video-ad-analyzer` for cuts, action, speech evidence and the relationships that drive the screen. Deliver analysis alone when requested. For a new script/product, create a separate semantic plan and pass it to the selected model writer; Seedance 2.5 uses `sd25-pe`. Original timestamps remain evidence, not fixed timing for rewritten dialogue.
2. **Previs or motion control:** Use the catalog's Blender workflow for reference analysis, primitive blocking, camera/action control, render review, and comparison with generated video. If the request is analysis-only, deliver the analysis without assuming scene creation is authorized.
3. **Video prompt:** Use the explicitly selected video model's writer. Keep its reference identifiers, field order and output format. A shared brief can feed several requested model writers, but their final prompts remain separate.
4. **Still-image assets:** Use the image writer only when the user needs an image prompt or image work, such as a product sheet, background, keyframe or edit. A still image for a video project is still an image deliverable. Image-to-video prompting goes to the selected video model writer.
5. **Generation or submission:** Prompt skills supply instructions, not a generation service. Use a separate available execution capability only when the user's request includes that action. Preserve an established manual-submission workflow.
6. **Feedback:** Compare the reference, actual submitted inputs and result. Distinguish geometry/contact, motion/timing, identity/surfaces and appearance. Route a confirmed cause to the affected previs, image asset or prompt, then verify the changed part.
7. **Metadata cleanup:** Use `video-scrub` when requested to inspect or remove identifying metadata from a local MP4/MOV. This is an independent file-processing stage, not a model prompt or shot analysis. Preserve the original and follow the skill's codec/HDR/caption and verification limits. Do not add cleanup automatically to every production task.

Use the user's model choice and prior decisions. If no model has been chosen, continue independent reference analysis, asset planning or previs work. Ask for a model only when a model-specific deliverable depends on that choice; do not silently choose one or force every project through Blender.

Every final copyable prompt, for video or still images, is written in English regardless of the user's language or a requested prompt language. Spoken dialogue, lyrics and in-scene or rendered text are the exception: they stay in the language in which they must be heard or seen (supplied wording verbatim, newly drafted lines in the intended language), never translated into English because the prompt is English. Explanations, analysis and handoff notes may follow the user's language.

## Carry context between stages

Maintain a compact brief containing the output medium, subjects and object identities, reference roles, required action and outcome, shot boundaries, product states, sound/text requirements, file versions and approved deviations. For dialogue-led adaptations, carry speaker roles, exact new dialogue, phrase/action anchors and the distinction between observed source times and estimated target duration. Use this brief instead of reopening the whole conversation or loading unrelated model guides.

For reference-driven video, distinguish what the reference controls: composition, camera, motion, timing, anatomy, identity, product design, surfaces, wallpaper, style or sound. Resolve contradictions by preserving the user's current intent and stating any material assumption. An appearance sheet must not accidentally become a multi-object layout instruction.

Keep output settings and platform attachment mapping outside the copyable prompt unless the selected writer's format requires otherwise. An `@` alias in plain text is not proof that a platform attachment is bound; native subject tags and actual attachment chips have different roles.

## Scope and library boundaries

- The catalog lists the active skills included in the distribution.
- Read references progressively. Resolve their paths from the containing document and keep each skill's supporting files together. Source notes provide attribution; they are not additional runtime dependencies.
- For speech analysis, reuse a prepared shared or project-specific WhisperX runtime through `video-ad-analyzer`'s speech guide. Keep runtime/cache paths separate from the current project's inputs and results; choosing this stage does not require reinstalling an available environment.
- Do not create images, render Blender scenes, run paid generations or publish content solely because a later production stage may need them.
- This router uses the catalog's sibling skills when present. If copied alone, locate the named installed skill through the host's available skill list; if missing, report the missing capability rather than inventing its model syntax. A generic draft may still be delivered when useful and clearly labeled.
- Keep observed results, hypotheses and pending validation separate. An untested prompt or previs is not a successful model test.

Deliver the requested artifact and a concise handoff: the relevant input files, exact prompt if requested, settings or bindings the operator must supply, and any unresolved issue that changes the result.