crazyooo Voiceover Visual Video Producer
Create a finished video whose continuous narration is carried by intentional main visuals. The skill is portable: do not assume a particular workbench, editor, project API, or directory layout already exists.
Establish the production
Use user-facing language first. Ask what material the user already has only when it is not evident:
- Script: pasted text or an uploaded text/document file. Generate narration with a configured MiniMax voice.
- Existing narration: an audio file, or a video whose audio is the intended narration. Extract or normalize the audio, then transcribe and align it.
- Subtitle file: an advanced optional input such as SRT, VTT, or ASS. Use matching supplied narration when present; otherwise generate narration with MiniMax.
Do not require the user to know what SRT means. Internally normalize every route to one narration master and one canonical subtitle timeline. Determine platform, aspect ratio, language, source restrictions, and delivery format from the request and available files. Infer ordinary defaults when safe: use 1920x1080, 30 fps, and 16:9 when no platform or source indicates otherwise.
Read references/narration-inputs.md before creating the timeline. The narration master must be established before visual production:
- For existing audio or video, use the actual narration as timing authority and create the internal subtitle timeline by transcription and alignment. Do not make semantic deletions unless the user requests narration cleanup.
- For text, require a configured MiniMax model and voice, generate a short sample when the voice has not been approved, then synthesize the full narration.
- For a subtitle file, validate its cues and pair it with supplied narration or MiniMax generation.
- Measure actual audio and create the canonical internal subtitle timeline from real speech timing. Text length or requested timestamps are not proof of duration.
MiniMax generation may incur cost. Apply MiniMax authorization rules only when generating voice: do not send a paid request without the user's explicit authorization, do not expose credentials, and do not automatically retry a failed or cancelled generation.
When MiniMax is required, check whether a usable MiniMax speech configuration is already available before asking the user to choose a voice. If credentials, model, or voice configuration is missing or invalid, read references/minimax-setup.md and guide the user through the available host's secure configuration path. Do not ask the user to paste an API key into ordinary chat or write it into project files. Verify configuration before offering a paid voice sample. If the user does not want to configure MiniMax, offer the existing-narration route instead of silently switching providers.
If brand materials exist, inherit them. Otherwise offer or select a preset from references/presets.md. Treat presets as adaptable visual systems rather than interchangeable color skins.
Production order
For end-to-end work, read references/workflow.md. Lock the narration master and canonical subtitle timeline before finalizing animation, screen recordings, captions, music, or edit points. Do not route this workflow through raw-video or A-roll rough cutting.
Build a coverage map for the entire narration. Every interval must have an intentional main visual: animation, diagram, real interface, image sequence, licensed footage, generated scene, or deliberate designed hold. Picture-in-picture and decorative overlays do not count as full-frame coverage.
Choose treatments by meaning:
- Use real interfaces, products, documents, and evidence when the narration makes factual or demonstrative claims.
- Use diagrams and Motion Graphics for systems, relationships, comparisons, quantities, and processes.
- Use imagery or footage for place, atmosphere, people, and concrete events.
- Use restrained typographic scenes for key claims or transitions, not as continuous karaoke text.
- Avoid generic stock montage that has only a thematic association with the words.
Record source and rights information for external assets. Never present generated imagery, mock interfaces, or decorative data as documentary evidence.
Default to a configuration checkpoint when MiniMax is unavailable, a voice-sample checkpoint when MiniMax is used, then checkpoints after the storyboard and representative visual sample. If the user explicitly requests uninterrupted end-to-end execution, continue without routine pauses, but still stop for missing credentials, a required missing input, external spending, rights questions, and destructive actions.
Truthful completion
Use supplied narration or the configured MiniMax voice service, plus available transcription, alignment, media inspection, graphics, image, audio, editing, and rendering capabilities. Follow the instructions of any specialized tool or skill actually used. Never claim a voice generation, transcription, render, source check, synchronization, or viewing pass that did not occur.
Before delivery, read and apply references/quality-gates.md. Deliver the playable final file when rendering is available, plus production artifacts needed for revision. If a capability is unavailable, preserve completed artifacts, state the exact blocker, and do not call a storyboard or animatic the final video.