I wrote a song about vibe coding (and built the whole music video with AI agents)

A developer built a full music video for a song about 'vibe coding' using AI agents, with nearly the entire production handled autonomously by Claude Code while the developer was away. The pipeline included AI-generated music, Whisper-based transcription for shot mapping, image-to-video character clips, and programmatic rendering of screen gags, all assembled via a Python EDL driving ffmpeg. The developer shared lessons on consistency, transcription, and trusting rendered pixels over API claims.

I used to write code. Now I write prompts. Then I review the code I didn't write. Then I fix the bugs the AI created. At some point this year that stopped being a complaint and started being... a chorus. So I made it one - a full 6-minute song about being a developer in 2026: It has verses about vibe coding, the AI subscription stack "my wallet says please stop, my brain says one more tool" , and the day I gave my coding agent one tiny task and it opened seventeen pull requests. The part you'll actually find interesting: the video is agent-built too I decided the production itself should prove the song's point. Almost the entire music video was produced by an AI agent Claude Code working autonomously while I was out: - The song — generated from my lyric sheet fun discovery: the generator silently dropped my spoken outro and invented a reprise instead; nobody reads the spec, not even the music AI . - The shot map — the agent transcribed the finished track with Whisper, cross-checked section boundaries against an RMS energy profile, and turned it into a 60-segment edit decision list with real timecodes. - The character clips — 16 scenes from an image-to-video model, kept consistent with a 4-image reference kit full-body sheet, smiling portrait, deadpan portrait, set reference . Every one was approved on the first take, which never happens. - The screen gags — the fake PR list with 17 open PRs, the "you're almost out of tokens" modal, the receipt wall — all rendered programmatically PIL frames piped to ffmpeg because you can't screen-record a browser while you're not home. - The assembly — a Python EDL driving ffmpeg: per-segment cuts, concat, the song muxed as the timeline master. Total human contribution to the edit: reviewing takes, catching alignment bugs, and one deeply 2026 code review comment: "the receipts are crooked." Things I learned that transfer to real work - Reference kits beat prompts for consistency. Sixteen video generations with the same face only worked because every prompt attached the same four reference images. Same lesson as agent context engineering: don't describe, attach. - Transcribe your media before editing it. Whisper timestamps turned a vibes-based edit into a deterministic script. The whole cut is reproducible from one Python file. - Fonts lie. PIL's glyph probe said Helvetica had ✓ and ✗. It rendered tofu boxes. Trust rendered pixels, not APIs — which is basically the song's thesis, applied to typography. The question the song ends on The tests are green. I don't know how it works. But hey — the tests are green. Which verse is your life right now? For me it's the subscriptions one. My credit card knows all the Pro plans by heart. Full song: https://youtu.be/TLDh- AiZVk https://youtu.be/TLDh- AiZVk — and if you want the production pipeline details, the tools are a few hundred lines of Python + ffmpeg; happy to write that up if there's interest.