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, natural light, handheld imperfection) timed to the lyrics, arranged as a 3-act arc (its shape is the user's story_arc choice) 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 (Sundrift "Roll On", ~28s 1080×1920 9:16, 14 tableaux) — copy its structure, never its creative values; 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-gpt-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 spend nothing.

This skill ships no assembler script of its own. Do the cuts and captions with the stitch-videos-ffmpeg skill's montage.py (../stitch-videos-ffmpeg/SKILL.md) and the anthem mux with one ffmpeg pass; scripts/PIPELINE.md step 4 lists each step.

Choices

The creative content this stage assembles is decided upstream by the format's choices, asked of the user before any paid step. The worked example's values are examples, never defaults:

• song_style — the anthem's genre, mood, BPM and arrangement. The demo used a sunny 120-BPM indie surf-pop anthem.
• vocalist — who sings it. The demo used a bright, warm female lead.
• cast — who is on screen in the tableaux. The demo used four young women roller skaters.
• setting — place + time of day (drives the look pack's light and palette). The demo used a seaside boardwalk town in late summer, golden hour.
• story_arc — the 3-act shape of the tableaux and lyrics. The demo used one summer day, morning → skate-bowl peak → twilight, with an origin-story wink.

This assembly reads them only through the config (tableaux[] windows + captions, the anthem, captions.accent_words, end_card.*); it hardcodes none of them.

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 (low on the frame) with the hook line and accent words in the accent colour. 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 captions, overlay the end card → a 1080×1920 h264+aac master. No paid calls, no keys.