Teach a coding agent to analyze a song's tempo, structure and parts

Developer Mager plays piano casually and had a 37 second clip of himself playing a pop song. He had Gemini draft a skill on the Loooom catalog for identifying song parts, published it through the web form, and asked Claude Code to analyze the MP3. Claude installed ffmpeg and librosa by itself mid session, wrote a Python audio pipeline and returned a report. It estimated tempo near 91 BPM, against about 104 for the original song, called the key B major, and mapped intro, verse, chorus and outro with timestamps. It also guessed which instrument layers were present. He notes three different AIs were involved and is still working out practical uses, such as practice feedback.