Claude Code hit 67.5% key match, missed the chorus

The author's experiment grew out of three motivations: a song with no published chords that they wanted to arrange, curiosity about Claude's limits, and a desire to skip the tedious parts of transcription — writing out the main melody and chords — so they could get to the fun part, harmonizing. They fed an mp3 to Claude Code, which used librosa for tempo and chroma analysis, Demucs for source separation, basic-pitch for transcription, and MuseScore to produce a score. The key estimate of 67.5% was measured against a 58.3% random baseline, a technique the author found broadly useful for reading any metric. The first chord attempt produced a chorus stuck on a single chord, which the author recognized as implausible even without perfect pitch. Claude then measured whether pitch information survived in the mix and found it largely absent, prompting a switch from improving inference to improving input. After Demucs isolation, the song's half-step modulation appeared, matching what the author had felt but could not name. A separate workflow issue — telling Claude the goal was an a cappella arrangement — caused it to filter out notes below vocal range, leaving only 23 bass notes; moving that constraint from input to output raised the count to 190. The author concludes the real gain was not speed but a shift from copying to verifying, work that felt less like a grind.