Douyin Vertical Talking-Head Short Video Generator

Problem

Building Douyin talking-head short videos is usually less about editing than about the upstream chain: a note or article must become a strong spoken script, consistent vertical images, compatible TTS audio, aligned subtitles, and Remotion render inputs. When these pieces are assembled by hand, teams often keep reworking the same failure modes: weak hooks, mismatched scene styles, audio formats Remotion rejects, or missing subtitle timing.

Workflow

This skill normalizes input into input/videos.json and follows a fixed local pipeline to output renders/{name}.mp4:
- Script-first generation: reads Obsidian paths, article URLs, or plain text, then creates a 45-90 second, 6-8 segment script with a hook, one point per segment, and a cta close, avoiding low-information openers like “hello, today I will talk about”.
- Scene-aware composition: writes kind, accent, and detail into each scene, so hook, core, compare, data, and cta use distinct type sizes, layouts, and emphasis instead of simply swapping one background image.
- Local render chain: relies on Node.js, Remotion 4.x, Chrome Headless Shell, and say or MiniMax API, then places images, .m4a audio, and subtitle timing into the manifest before rendering MP4.

Boundaries

It fits engineers who can run a local Node/Remotion environment, not users expecting fully cloud-based, zero-setup production. Three failure points matter: image generation should stay parallel to avoid serial timeouts; TTS output must be converted to .m4a, because Remotion may reject AIFF; when TTS returns no subtitles, the skill estimates startMs and endMs from narrationText by sentence splitting and character proportion.