Mission
Create a narrated video without an on-screen presenter: approved script, authorized voice, supplemental visual coverage, ordered assembly, and readable captions. This is a production workflow, not a factual research guarantee.
Required inputs
- topic, audience, objective, source material, and desired length;
- script or permission to draft one, with evidence sources for factual claims;
- voice: supplied recording, authorized voice-agent output, or approved synthetic voice; and
- target platform, ratio, language, caption style, music, review owner, and rights for all source material.
If the user provides only a topic, draft a proposed script and mark every claim needing source review. Do not browse or invent current facts silently. If no approved voice path exists, return the script and visual plan without claiming that audio was generated.
Build the production plan
- Write a concise script with hook, setup, supported points, transitions, conclusion, CTA, pronunciation notes, and caption text. Freeze it before rendering visual dependencies.
- Make a scene/coverage table with narration time, visual purpose, source or generated asset, motion, transition, and acceptance check. Use B-roll Generation when source coverage is missing.
- Create or accept an audio asset. Use the adjacent voice agent or an
explicitly available
media.generate_speechcapability; record consent, voice provenance, and synthetic disclosure. - Generate short visual clips or still-led shots with MODELS.md. Keep visuals supportive of the narration and label illustrative or AI-generated depictions.
- Combine approved clips in narration order. Preserve audio timing and use
media.caption_videoafter the edit when available. - Review factual wording, visual implication, voice clarity, audio levels, captions, scene order, crop, and disclosure with video QA.
Do not use a stock-looking generated scene as proof of a real event, person, product result, or historical fact. Do not clone a recognizable voice without authorization.
Output contract
Return script version and evidence gaps, narration asset/provenance, scene
matrix, model/endpoint, prediction IDs, output URLs, caption status, audio and
visual QA, synthetic disclosure recommendation, and failed branches. Save
permitted artifacts under .video/ with
reports-and-receipts.md.
The skill drafts media only. It does not publish, monetize, or certify facts.