AI Short Video Creator

Problem

When clips, narration copy, and subtitle timing are scattered across multiple files, manually aligning cut points, generating speech, proofreading captions, and adding transitions is error-prone. This skill separates the workflow into two stages: first produce reviewable narration and subtitle materials, then assemble the video only after explicit approval, reducing hard-to-fix rendering mistakes.

How It Works

The pipeline is organized around a SCRIPT data structure. Each narration segment is mapped to one or more source clips, with total clip duration slightly longer than the generated voice-over. In the first phase, it uses ffprobe to build a clip inventory, calls edge-tts to create audio, and extracts subtitle timing from SentenceBoundary events to write subtitle_review.md. It then presents the clip inventory, segment scripts, subtitle table, and segment-to-clip mapping so a user can revise text, timing, clip order, or substitutions before proceeding.
In the second phase, it assembles the video: trim clips, scale them, concatenate segments, burn subtitles using an .srt file and the FFmpeg subtitles filter, add xfade transitions, merge the AAC narration, and export an approximately 1080p, 30 FPS H.264 video. It fits Chinese voice-over and short explanatory videos. If source material is insufficient, the narration is too long, file paths are unusual, or complex multi-track audio is required, the timeline needs manual review.