Harden video autopilot production workflow

[danne135]

What changed

• Added a portable skills/video-autopilot/ Codex skill specification.
• Added repeatable rules for measuring storyboard grids, correcting 9:16 framing, selecting YT_music / ACE-Step candidates, confirming narration, and mixing voice with BGM.
• Added formal MP4 and audio QA requirements to the public workflow documentation in both Chinese and English entry points.

Why

The succulent teaching-video run exposed three reusable failure modes: incorrect assumed grid coordinates caused neighbouring-panel slivers, procedural music sounded rigid, and narration needed a confirmed audition plus per-scene timing.

Impact

Future image-first teaching videos can reuse the same production contract, preserve intermediate evidence, and hand off a versioned MP4 with auditable crop, audio, and subtitle decisions.

Root cause captured

The source storyboard grid was not the assumed size. The skill now requires measuring the actual image and separator pixels before panel extraction, then checking a contact sheet from the formal MP4.

Validation

• skill-creator validation: Skill is valid!
• git diff --check: passed.
• Local demonstration video: 1080×1920, 60 seconds, AAC 48 kHz; audio_qa.py passed at -16.3 LUFS, -2.1 dBFS true peak, and 0 seconds longest silence.

[Hao0321]

Adoption review: The proposed workflow uses a second skills/video-autopilot location instead of the published codex-skill/video-autopilot entry, names an audio_qa.py/YT_music workflow that is not provided by this source tree, and requires portrait QA across a tool that supports landscape work. Useful framing/audio checks need an integrated public implementation, correct routing, and same-material output comparison before an improvement claim. No claim of malicious intent is made; no approval or merge is given.