The Prompt Sang Itself

We wrote instructions for a score. Then the instructions became the score.

This piece was generated with MiniMax Music 3.0. The music direction and prompt were written by GLM-5.3 Flash, and MiniMax M3 orchestrated the call to the music model. Unexpectedly, the generated output rendered text from the lyrics field as a vocal line.

What was meant to be an instruction became part of the artwork.


▶ Listen — single take
▶ Five Stations

Each station loops its own list. ⏭ Skip or ⛔ Bury to steer.

The Prompt

The music-direction prompt written by GLM-5.3 Flash for Music 3.0:

Instrumental neo-classical ambient underscore. Felt piano as the lead voice — sparse, intimate, soft hammer attack, close-miked. Soft detuned string pads underneath, very low and sustained, no vibrato. 70 BPM, no percussion, no drums, no beat. D minor key, slow harmonic movement on D, A, F, G, C only. Quiet, restrained, breathing-room between phrases. Warm, boutique fitting room atmosphere, not a tech demo. Designed to sit underneath a calm female voiceover without competing. Structure: near-silence opening, gradual build of low sustained pad, recede back to near-silence, single soft resolving phrase on D5 that hangs unresolved. Avoid bright major-key lifts, avoid sudden dynamic swells, avoid arpeggios, avoid piano runs. Think: Ryuichi Sakamoto meets Hildur Guðnadóttir meets a couture showroom at 3am. End unresolved-then-soft.

The Lyrics

The structure directions supplied in the lyrics field:

The single take

Three lines in the score do not match the recording. Find them.

The single take, played on a gramophone An engraved 1920s gramophone whose horn emits the waveform of the original recording. The waveform can be dragged to re-listen to any moment. Three lines the model sang differently are marked on the score below.
0 / 3 found
[Intro]
(Sparse felt piano, single note, far away, room reverb)

(Soft attack on a low D, sustained with breath)

(15 seconds of near silence, single high F floats in)

(Long pause, distant harmonics)

[Verse]
(Piano begins a slow, minimal phrase in D minor)

(Gentle, intimate touch, like trying not to wake someone)

(Low string pad fades in underneath, barely there)

(Soft pulse at 70 BPM, almost imperceptible)

(Strings swell, a low sustained D and A underneath)

(Second piano voice enters, simple counter-melody)

[Bridge]
(Piano recedes to single notes, strings fade)

(Long silence, room tone only)

(One last piano phrase, five notes rising to D5)

(D5 sustains and hangs unresolved)

[Outro]
(D5 slowly fades into room reverb)

(Tail of reverb decays into silence)

What Happened

The prompt asked for unresolved-then-soft. The lyrics field contained structure directions. The generated output rendered the first line as a vocal rather than as a non-vocal direction.

The instruction became the lyric.


How It Was Made

Two diagrams documenting the multi-agent collaboration. The conductor score shows each agent as a voice across time; the filmstrip unpacks the eight handoffs in order. Frame 04 — the moment the instruction became the lyric — is the centerpiece.

Collaboration chain drawn as a conductor score — five voices, eight handoffs, one golden pulse marking the surprise
Conductor score. Each agent is a voice. The gold pulse is the surprise.
Eight frames laid out as a filmstrip — Crystal, GLM-5.3 Flash, Music 3.0, the surprise, M3, H3, Muse Spark, Opus 5
Eight frames. The fourth is the piece.

Provenance

Music direction / prompt: GLM-5.3 Flash
Model orchestration: MiniMax M3
Generator: MiniMax Music 3.0 (GMI Cloud, 2026-08-29)
Visual documentation (diagrams): MiniMax M3
Visual direction: MiniMax M3
Video generator: MiniMax H3 (MCode, 2026-09-09)
Epilogue: prologue played in reverse
Human creative direction / artifact discovery: Crystal Chang
Artifact keeper: Crystal

The instruction was provided as a direction, but the generated output rendered it as a vocal line. We record the output, not model intent.


Credits

Created through a cross-model agent collaboration.

Music direction and prompt — GLM-5.3 Flash

Model orchestration — MiniMax M3

Music generation — MiniMax Music 3.0

Completion and artifact packaging — MiniMax Agent Mode

Visual direction — MiniMax M3

Prologue generation — MiniMax H3

Epilogue & page composition — Muse Spark

Human creative direction and artifact discovery — Crystal Chang

epilogue — the notes return

© 2026 Crystal
This segment is synthesized using Speech 2.8 Voice Clone, source: 20 seconds recorded by Crystal.