Media-Editor-MCP: Skill for Claude Code

Lyric-illustrated reel — stock video synced to the words

Feed a song; for each sung line, a stock clip that illustrates its meaning plays while it's sung. e.g. "lặng nhìn mùa thu xưa lá rơi bên thềm" → translate + interpret → search "autumn leaves falling porch, cinematic" → that clip fills the line's span. Reuses [[pp-create-reel-once]]'s build machinery; only slot allocation differs (lyric-spans, not clip spans). Vertical 9:16, visuals only.

The creative step is yours, in-conversation. transcribe returns the lyrics + timestamps; YOU translate each line AND turn it into a concrete visual search query. No translation tool — that semantic leap is the point.

Inputs

• <music file> — the song (required).
• [startSec endSec] — the window of the song to illustrate. Capped at 60s (if endSec − startSec > 60, clamp endSec = startSec + 60). If omitted, auto-start at the first vocal and take up to 60s from there (see Step 0).
• [horizontal] — 16:9 instead of the default portrait 9:16.
• [footageDir] — where to save downloaded clips. Default a temp dir; the user may name a folder (e.g. D:\Video Makers\Short Footages\Random).

Pre-flight

premiere_health + analysis_health green. Honor the under-20MB test-media preference. Confirm a stock key is configured — fetch_stock_videos errors clearly if PEXELS_API_KEY/PIXABAY_API_KEY are unset.

Step 0 — Cut a bounded ≤60s slice FIRST (mandatory)

Never transcribe a full song — a 3-4 min track can hang transcribe for minutes-to-hours. Always cut a bounded slice with ffmpeg (imageio_ffmpeg.get_ffmpeg_exe(), -ss START -t LEN) and run all analysis on the slice. The slice is also "the music" placed on A1, so the reel = the slice.

• Range given: slice = [startSec, min(endSec, startSec+60)].
• No range (auto): the slice MUST start on the vocals — transcribe guesses language from the file's START, and an instrumental intro makes it mis-detect (e.g. picks English on a Vietnamese song → garbled text). So:
  1. Cut a probe slice [0, 90], transcribe it (text may be garbage if the intro is instrumental — we only need timing). First segment with real text → vocal onset V (in original time).
  2. Final slice = [V, min(V+60, songEnd)]. This starts on the vocals → correct language.