Render Cinematic Music Video Agent Skill

render-cinematic-music-video

Assemble a cinematic music-video ad from a config: a live-action-STYLE short film where an

original sung anthem is the score and every visual beat is a shot-on-film tableau (Kodak Portra

grain, light leaks, golden hour, handheld imperfection) timed to the lyrics, arranged as a

3-act arc (morning → peak → twilight) with the hook line on the chorus drop. This capability is

the FREE, deterministic assembly — cut-to-window, hard-concat, anthem mux, caption burn,

and the brand end card.

scripts/config.example.json is the worked example (Hype and Vice "Game Day Girls", ~28s

1080×1920 9:16, 14 tableaux); scripts/PIPELINE.md maps every config block to its source step

and scripts/README.md documents the free assembly.

Run

This is the FREE, deterministic assembly stage — it spends nothing. The paid inputs are

separate capabilities: the sung anthem (create-music-elevenlabs, force_instrumental false —

the lyrics ARE the script, returns mp3 + words_timestamps); one 35mm-film keyframe per beat

in one look pack (create-image-fal); and one Kling 3.0 i2v clip per beat (create-video-fal).

Given the delivered anthem + words.json + one clip per beat + the brand end-card asset,

render-cinematic-music-video cuts each clip to its lyric window, hard-concats on the beat,

muxes the anthem, burns the cinematic lower-third captions, and overlays the end card → the

master. Re-cuts reuse the existing anthem / keyframes / clips and cost $0.

Contract (the free assembly)

• The sung anthem carries the story — no separate VO. The generated ElevenLabs track IS

the bed and the script (force_instrumental false); do not add a spoken voiceover or a

second bed.

• Plan the timeline AROUND the delivered anthem. The anthem is generated first and

reshapes/overshoots length; snap every tableau boundary to the lyric-phrase edges in the

returned word timings — never trim the anthem to a pre-planned grid.

• Captions from the anthem's OWN word timings, not Whisper. Derive words.json from the

music model's words_timestamps, chunk ~4 words at lyric boundaries, and burn cinematic

lower-third serif captions (--placement low) with the hook line accent-treated (bold-italic).

Whisper on sung audio returns "🎵 Music Playing 🎵".

• Land the hook on the chorus drop. The hero tableau (is_hook) is timed so the

load-bearing line sits on the chorus drop; accent that line in the captions.

• One cinematic look pack + hard cuts on the beat. The look pack (named film stock + grain +

flares + handheld) plus a continuity anchor holds every clip together; cut each clip to its

lyric window and hard-concat (one optional match-cut into the hero reveal) — no dissolves.

• End card from the real brand asset — never AI-render the wordmark as diffusion text.

Composite the lockup (PIL/ffmpeg) from the real asset or use a designed keyframe base;

diffusion garbles a wordmark.

• FFmpeg composite, deterministic, FREE. Normalize each clip to its beat-locked window,

hard-concat, mux the anthem (afade in/out + loudnorm I=-14), burn the caption ASS, overlay

the end card → a 1080×1920 h264+aac master. No paid calls, no keys.