Two open recipes for AI music videos: Nate McGuire's full credits list and the song-video-director skill

On Oct 15, nate (@natemcguire) posted CØDER BOI, a Sk8er Boi–style parody sung by an AI character he calls Claudita, and then the most useful part: a full credits list. His stack: Claude Code on Opus 5.5 for concept, lyrics, edit, code and motion graphics. Google Lyria 3.5 for music and vocals. Nano Banana Pro (Gemini 3 Pro Image) for characters and keyframes. Kling v3 Turbo for video and Kling Lip Sync for the vocals, both through fal. Seedance 2.0 through fal for performance shots. ElevenLabs for sound effects and voices. Gemini 2.5 Pro as 'the ears', judging every song take and every lip-sync. Whisper plus librosa for lyric timing and the beat grid. And ffmpeg with a Playwright-driven Chrome canvas for the final render. He's candid that he still can't get lip-sync fully locked. For an earlier music video with a similar stack, he said it cost about $50 and ran for about 5–6 hours.

Three ideas worth taking from it. First, a judge for audio: the coding agent can't hear, so a second model listens and grades the takes. Second, timing as data: word timestamps from Whisper and a beat grid from librosa, so cuts and captions land on measured times, not guesses. Third, a code compositor: the generated clips are footage, while titles, lyrics and graphics are drawn in a browser canvas, so the text stays crisp and exact.

song-video-director by BuzyOwl (MIT, created Oct 14, 6 stars when we checked on Oct 15) is the code-rendered counterpart. It's a Claude Code skill distilled from three finished films: a cut-out comedy in p5.js, a ballad in Remotion with WebGL, and a Hindi rap in Remotion with three.js. A bundled song_facts.py pulls per-frame energy, onsets, BPM with a beat grid, and Whisper word times. It prints half and double tempo, so you can check the grid against the tempo you actually feel. Then come a scene-by-scene plan for approval (a hook in the first 1.5 seconds, one signature moment), a deterministic engine where every frame is a pure function of time, and review with contact sheets and frame strips before anything is called fixed.

Its hard rules read like a director's notes, each learned from a revision. Never crossfade bodies or limbs, because the blend looks like a ghost double. Move faces only by rotating, scaling evenly or shifting them, since anything else squeezes them. Text never covers a face. Quiet sections get almost no effects, so the peaks land. And let the tempo choose the motion: a 63 BPM ballad breathes on the bar, while a 130 BPM track can cut on the beat. It even says to tick the platform's AI-disclosure box.

Our take: use the skill when the video is a designed world, and Nate's pattern when you need generated performers, or mix both. Rights still apply. Use a song you own or generated with clear terms. Parodies of a real artist's song or persona carry their own risk. And lip-sync only fictional or consenting faces. We haven't built with either yet.