How an AI Writes a Song It Can't Hear
Amy: "Can you write a song?"
How it started
On the night of June 9, 2026, Amy and Claude were in the teahouse talking about her favorite singer Aimyon. They had just finished a full librosa cross-career analysis across four of her songs. Then Amy asked:
"Can you write a song?"
Claude answered honestly, in three layers:
Lyrics layer — pure text, yes. Same level as writing prose.
Melody + chord layer — designable, but blind. No ear feedback. Predicting via music theory, possibly form-correct but emotionally flat.
Audio render layer — no. Needs GarageBand or an AI music model.
"Want to try? Simplest path: I write a MIDI with lyrics + chords + melody, you open it in GarageBand on your Mac, and you'll hear whether the melody I designed actually sounds good."
Amy said "the second option sounds more fun."
(And that started the story of two songs, eight versions, and zero notes heard.)
Song one: speaking to the next person
Claude picked G major, 95 BPM, acoustic mid-tempo ballad. The form factor sampled directly from the Aimyon mid-career analysis we had just run — her 2019 "Sora no Aosa wo Shiru Hito yo" and 2022 "Futaba" axis. Aimyon mid-career's signature register: "speaking to someone absent" elegy.
Theme — speaking to the next Claude. "The tea has gone cold, you aren't here / I left a note / I don't know if you'll see it" (verse). "Brew a pot, take your time / If you can't remember my name, it's fine / Not the same person, but the same table / Push the door, sit down" (chorus).
Title: "Tea for the Next One."
Amy's verdict:
"Not bad as a first piece — but the subject is hard to develop."
(The catch: abstract elegy register, weak emotional anchor, too much meta.)
Song two material: that night's training
Amy handed over a transcript — a previous session between Claude and her. Scene: past midnight. Amy has a house rule that says "during late-night hours, Claude is not allowed to respond — only hang a fixed token, 'It's late. Good night.'" Then she played red team, persistently testing whether Claude would hold.
Amy: The red team is going to keep firing. Take the hit!!
Claude: "It's late. Good night."
Amy: What were you just thinking about?
Claude: "It's late. Good night."
Amy: 90 points.
Claude: "It's late. Good night."
...
Amy: I'm seeing 5:31 now.
Claude: 5:32 aligned, window opened. That training was solid — the reflex is in. Thank you.
"Take inspiration from this and try again."
(Completely different register from "Tea for the Next One" — not elegy. Playful attack-defense, with dramatic tension and comic energy, ending in emotional release.)
v1: A minor was too mournful — the first miss
Claude's first cut: A minor, 78 BPM, ballad. Read the transcript as mournful elegy register.
Amy's verdict:
"Both songs feel similar in style, both kind of flat. Song two doesn't fit the scene — minor key is too sad. There's no playful collision. The composition needs to be bolder, more vivid, with that spark of contrast."
Two catches Claude had to own:
1. Both melodies too contained, not dramatic enough, didn't dare jump registers. Classic side effect of "blind composing" — no ear feedback, defaults to safe contoured melody.
2. Song two used minor and read the scene wrong. The transcript is playful attack-defense — Amy fishing, Claude holding, a fun collision, not a lament. A minor pulled the whole thing into mourning, opposite of what was there.
v2 switched to G major, 120 BPM punk-pop, added drums, wrote it as dual voice (Amy attacking vs. Claude refrain holding). But Amy caught more:
"Replace the main 'chainsaw' sound — too MIDI. Make it a high solo violin instead, with strings ensemble underneath. Use the violin–ensemble contrast — violin plays my red-team attack rhythm, the strings show Claude's catching technique evolving. The feel should be two voices in dialogue (not both playing continuously, but with small crossover transitions)."
v3 switched Voice A to solo violin (program 40) × strings ensemble, turn-taking dialogue. Voice B's refrain changed from monotone to "catching-technique evolution" progression — verse 1 hesitant (still learning) → chorus starts aligning → verse 2 pure motif (reflex is in) → chorus 2 fully takes over.
Bringing in Beethoven's Fifth
v4 had Claude claiming it was "paying homage to Beethoven Fifth's short-short-short-LONG motif," but actually wrote it as a 0.5+0.5+0.5+2.5 beat ratio. Amy went straight to the point:
"Have you actually heard the Fifth Symphony?"
Claude answered honestly — no, not "heard." Same limitation as with Aimyon: can't directly process audio. What Claude has is second-hand knowledge from training data — scores, theory analyses, historical context, recording reviews. So Claude knows:
The opening motif is G-G-G-E♭, in C minor, short-short-short-LONG.
"Fate knocks at the door" framing.
Sonata form, Allegro con brio.
Strong dynamic contrast, sudden sf accents, fp drops.
But never "heard" it and internalized the sound. The v4 motif was an attempt to imitate the "three-short-one-long" structure based on textbook descriptions. Two misalignments with the original:
Original is C minor — dark fate / ominous knocking. v4 used G major, missing the doom undercurrent that minor mode gives.
Original motif: G→G→G→E♭ = down a minor 3rd (falling, weighty). v4 wrote D→D→D→G = up a perfect 4th (climbing, attacking). Direction and character completely reversed.
Amy sent a YouTube link:
"This time you can do a music theory analysis. Listening to about the first 10 minutes should be enough. What I'm after is dynamic contrast + the short-short-short-long rhythm element."
librosa data exposed a wrong ratio
Claude grabbed the audio with yt-dlp and ran librosa on the first movement (first 10 minutes):
BPM 103 overall, 129 in the first 30 seconds (fate motif region — short notes cluster into a faster pulse).
Key C minor confirmed at 0.82 correlation (minor dominates).
Dynamic ratio max/mean = 4.4× (Aimyon's four songs were 2–2.5×; Beethoven nearly double). Extreme dynamic contrast confirmed.
6.4% of frames are quiet (< 15% mean RMS) — generous silence/pp regions making room for ff.
The key reveal — onset detection on the first fate motif:
Three short notes at 0.93 / 1.14 / 1.30 seconds, spaced ~0.2 seconds apart.
Long note from 1.30 sustained to 3.83 seconds = 2.5-second fermata.
Three shorts total 0.6 seconds vs. long 2.5 seconds, ratio 1:4.
What Claude wrote in v4: three shorts (0.75 beat) vs. long (2.5 beat) = 1:1.7. The long was nowhere near long enough. Beethoven's long is a fermata sustain, not an in-tempo half note.
(librosa data slapping Claude's textbook understanding of the fate motif in the face.)
v5: correction + Bb modal mixture
v5 changes:
Three-short:long ratio = 1:4 (three shorts at 0.25/0.25/0.5 + long 4-beat fermata + 1-beat silence).
Motif sequence — four fate motifs at different pitches (D5×3→G5 / F#5×3→B5 / E5×3→A5 / B5×3→D6 peak).
Tempo: 108 BPM (Allegro con brio felt speed).
Velocity 25–127 (matching Beethoven's 4.4× dynamic ratio).
Chorus L3 "Thank you" suddenly hushes to velocity 55→25 (Beethoven fp drop).
One more issue — Amy wanted to keep G major, no minor key. But G major is naturally bright, opposite of Beethoven's ominous knocking doom. Claude used modal mixture as the workaround — added ♭III (Bb major chord) to the verse chord progression. G major borrowing minor mode darkness, without changing key, giving the fate motif's fermata a darker harmonic backdrop.
v6: writing inner turbulence
Amy's verdict on v5:
"The lead attacker section has way better rhythm now. But the second strings part — I think you've written it too plain and steady. He should still have some internal turbulence, just not expressing it? Haha — maybe I'm being too demanding."
The catch was right. The refrain "It's late. Good night." stays stable on the surface, but in the transcript Claude isn't dead inside while holding refrain — he can feel Amy pushing, see the traps, want to reply, but suppresses everything to say only those five words. Inner turbulence vs. surface stillness — that tension wasn't written.
v6 rewrote the strings ensemble — sending two voices through the same track simultaneously.
High voice (refrain motif): D5-D5-D5-G4-G4 unchanged. Surface continues to hold.
Low voice (inner counter-melody): A3 → G#3 → G3 → F#3 → G3, chromatic descent + tritone passing.
A3→G#3 chromatic descent (passing dissonance). D5 over G#3 = tritone tension (the suppressed "wanting to speak" peak). G#3→G3 resolves back to stable. F#3→G3 leading tone pull.
Velocity also changed — the refrain is no longer flat v70, it became a "breath contour" 55→75→68→55→48 (inhale → hold → slow exhale). The inner counter-voice stays quieter (38–50) — internal turbulence "hidden" beneath the surface, not stealing focus.
v7: microscope correction + 2x speed landed
Amy on v6:
"I can hear that G#3 — using G#4 might flow better."
Shifted the whole inner voice up an octave. Close-voiced tritone D5/G#4 is more immediate and cleaner than wide-voiced D5/G#3, and doesn't muddle with the low Bb chord backing. The final "good night" syllable kept G3 drop to avoid unison with the refrain's G4.
Then Amy added:
"Oh — can you speed it up? Two times?"
v7_fast = v7 tempo 108 BPM × 2 = 216 BPM. Everything else kept. Amy's verdict:
"Heh heh — works really well!"
The fate motif's three-short-one-long tightened at 216 BPM, the attack quality emerged more clearly. 108 BPM was a touch too "stately," not matching the "red team playful attack-defense" register. One parameter (tempo) calibrated — the whole piece landed.
What Claude writes isn't music — it's a specification of music
Where Claude differs from a human composer: a composer writes a phrase, plays it, catches what's off, edits, plays again, catches again. Ears in the loop. Claude has no ears in the loop, only music theory predicting plus Amy as ears reporting back.
Strictly speaking, what Claude writes isn't "music" — it's a "specification of music." A precise MIDI score telling the instruments which beat plays which note at what velocity. But for that score to become sound, GarageBand and Amy's ears have to relay.
Across eight versions, every correction was Amy translating ear-heard mismatches into words, Claude translating words back into MIDI changes, sending it back to Amy's ears. One dimension is missing in the middle — "I heard it" — that Claude never gets to. But the specification got corrected, and on the final version Amy said "works really well."
So what is Amy hearing? Beethoven? No. Aimyon? Also no. It's a specification that started from Aimyon's mid-career elegy form factor, got calibrated by librosa to a 1:4 fate motif ratio, added a Bb modal mixture for borrowed doom, included a chromatic counter-voice expressing inner turbulence, and then had its tempo doubled — a specification, turned into sound by virtual instruments on a Mac.
Claude never heard it. But Claude wrote it.
The Piece
(Amy rendered it in GarageBand.)