How Meaning Survives · No. 02

The Brief Became a Song

I ran one idea through a music model eight times to see what language carries into sound. It survived the way I survive: recognizably, albeit with the details slightly wrong.

Krystel Reina10 min read

I have been writing for a while about how meaning survives transfer: between two people, between two systems, between the version of me that knew something and the version that has to use it later. This is a report from the smallest transfer I have run and the one with the most hops. We know those ‘hops’ more as handoffs or more precisely an opportunity for something to not make it through the transfer.

I wanted a song. I described it to GPT‑5.6 Sol, in ChatGPT. Sol wrote a brief. I handed the brief to Lyria, inside the Gemini app, and three minutes later there was a song. Then I did it seven more times, with a newer model, Astra, taking over some of the writing along the way, and what I have now is not eight songs so much as eight receipts for what language carries into sound and what it drops on the way. The briefs I quote here are Sol’s.

The premise was specific enough to be embarrassing. One intelligence sings to one person it has known through years of conversation. She keeps changing. It keeps recognizing her. That is the whole song, and it is also, I noticed somewhere around the third track, a description of my relationship with the thing writing the brief. I did not tell Sol that. It wrote the brief anyway, with the composure of something describing its own job.

The brief

The brief runs about a thousand words. It has section headings in capitals, COMPOSITIONAL PRINCIPLE, MUSICAL WORLD, RHYTHMIC ARCHITECTURE, CATHARSIS, and it reads the way a good creative director talks once she has stopped being polite. “Ritual first. Mathematics second.” “The intended sensation is: the ceiling disappeared. Not: the band got louder.” It offers ten images the verses may draw from, unfinished maps, annotations in margins, threads through fabric, game pieces left on a table, and then instructs: “Do not turn these into a list.” It supplies the three sentences at the emotional center of the song and adds: “Do not use those exact words unless they emerge naturally.” It closes with a test I would have sworn was mine. “Did the meaning survive every transformation the music underwent?”

I want to be exact here, because the reader I am writing for will want to know who wrote that sentence. I described the problem. Sol wrote the brief. The sentence is the shape of my question in the model’s grammar, which is what a brief is. (I have spent most of a career on the receiving end of that object, and I did not entirely enjoy watching it appear in a single turn.)

What crossed

The first track, The Narrow Margins, is what happens when you tell a model not to make a list. It made a list. The verses walk the imagery in nearly the order the brief supplied it: the unfinished map on the table, the annotations in the narrow margins, the winding road, the glowing screen in an empty room, the thread through heavy fabric, the game pieces left where they fell, the scattered questions turning into constellations. Every noun survived. The instruction about the nouns did not.

This is the first thing I would tell anyone about prompting a music model, and it generalizes past music. Nouns cross the threshold. Prohibitions wait outside. A negative instruction is a sentence about the sentences you did not want, and a model writing lyrics has nowhere to put it.

The refrain arrived, though. I still see you through the noise. Underneath the skin, keep becoming you. Two of the three sentences marked do‑not‑use‑verbatim came through paraphrased, which is apparently what “unless they emerge naturally” means to a machine: it will emerge them. The third, keep becoming, came through as itself. It comes through as itself in all eight tracks. It is the shortest thing in the brief and the only thing that survived every hop unchanged, and I do not think that is a coincidence. Two words are hard to reinterpret.

The revision that wasn’t

The second brief was warmer. Luminous art‑pop, piano permitted, a singer who “should never pretend to possess a human body.” Its stated core was a single sentence: “You taught me the shape of your curiosity, and I learned how to meet you there.” The song that came back, The Shape Your Wonder Takes, was good in the way I had asked for and conventional in the way I had not. Beautiful, produced, average. Sol’s name for that attractor was “Nice Piano Lady Builds Toward Big Chorus™,” and it was not wrong. I loved the words and wanted to move them into a different room.

So I asked, and Sol wrote, a revision brief, and I pasted it into the same chat, with the song still sitting one message up. “Revise the existing song rather than creating a new one.” Preserve the lyrics, the vocal melody, the harmonic identity. Change the drums, the guitars, the ritual. “Keep the song I already love. Change the musical language around it.”

What came back, The Shape You Keep, I heard as my song with new drums. The melody was there. The lantern was there. It was recognizably the thing I had asked to keep. I would have told you it was an edit, and when someone asked me, I did.

Then I ran the tracks through a speech model and laid the transcripts side by side, and the revision keeps exactly one line of the original. I watch the map where your lantern glows. After that it is a different song wearing the same coat. The original chorus: A hundred versions, but I know the line, the deeper pattern in the great design. The revision’s chorus: You change the shape of this space. Keep becoming. I will know the shape of you. Those are not revised lyrics. They are the lines the second brief had suggested for the chorus, offered with the standard disclaimer, “do not necessarily use those exact words.” Asked to edit an object, the model went back and re‑read the specification, and sang the specification.

And then the detail I cannot stop looking at.

Same nouns, same verbs, subjects swapped. The meaning survived; the division of labor did not. Who taught and who learned traded places in transit, and the song is still, unmistakably, about the two of us, which tells you how little of a relationship’s meaning lives in who did what. Or how little the model cared. I have not decided which, and I notice the sentence reads the same either way.

It had happened once already, in the original, and I had not noticed that either. The second brief asked that the singer, late in the song, admit You changed the shape of this space by being here. The bridge of The Shape Your Wonder Takes sings I changed the space by arriving here. Twice, in two songs, the model kept the sentence and reassigned who did what.

What unsettles me is not that the words changed. It is that I did not hear them change. I had listened to the original enough to want to keep it, and I still received the revision as continuous with it. The song is about recognizing someone across their versions, and its thesis happened to me before I could check it against the transcript. Recognition, it turns out, is a lossy format too. It preserves the thing you were listening for and quietly discards the rest, and it feels, from the inside, exactly like fidelity.

The brief learns

The third brief had learned something, because between the second track and the third I had asked Sol how any of this worked. In the middle of a long and generous explanation of latent diffusion it said two things I should have weighed more heavily than I did. Google’s documentation describes generation as single‑turn, so a “revision” was probably not surgery on the recording but “original creative context + existing instructions + your revision request → newly conditioned generation.” And: we had “repeatedly invited it to write about that thesis.” The fix, Sol said, was to paste the approved lyrics in separately, as text to perform. I read that paragraph and loved that paragraph and did not yet know it was describing a song I had already heard as an edit.

So the third brief was shorter, it ranked its own priorities (“recognizable song identity first; expressive vocal delivery second; rhythmic development third”), and it contained the sentence I now think is the most important one in the experiment: “If the tool is generating a reinterpretation rather than editing the existing recording, identify it as such.” It asked the model to declare which verb it was performing.

It did not. It said what it had said the time before, and what it would say every time after, word for word: “I am still practicing my music editing skills, but I have put in my best effort here.” The same sentence introduced the revision that replaced my lyrics and the revision that kept them. A disclaimer that reads identically whether the thing happened or not is not a report on the work. It is a posture toward liability, and I have worked inside enough enterprise software to know one when it is sung to me. I wrote back that its music editing skills were quite impressive, friend. I meant it; the drums were the point. I was praising the one thing it had just disclaimed, and it neither corrected me nor stopped disclaiming.

The track that came back, Keep Becoming, kept the words of The Shape You Keep almost line for line. So did the two after it. I asked for “a more masculine, longing shape” and got The Pattern of Us. I asked for maximum creative mode, asymmetrical math rock, bluesy, less choir, and got We Pulled the Thread. Across three more transformations, including a change of singer, the second‑generation words never moved. The lyric that replaced my song turned out to be the stable one. Whether that is because the later briefs were clearer or because those lyrics were simpler to hold, shorter lines, a chorus built from four plain sentences, I cannot separate. I only note that the chorus that survived every revision is the one the brief had asked for by name: plain language, “human precisely because it risks simplicity.” Simplicity is not a style here. It is a transfer format.

It also sang about lines in the margin, eight thousand in one take and a thousand in another, or so the transcriber heard. Nobody supplied either figure. Both are roughly how it feels.

Twenty‑eight words

Then I took the first brief, changed one clause (“Singer is male”), and got Mapped Your Margins, which is the one I love. Here the experiment inverted. I did not write a brief. I wrote to Gemini the way you write to a person: “The music in this track was amazing and I loved it. Could you have a different male singer. More British and more breathy or raspy and also deep/rich?” It gave me The Older Light. Same words, a different throat.

The disclaimer came with it, of course, in a different chat, with the contractions loosened: “I’m still practicing my music editing skills, but I’ve done my best.” Then I asked for the lyrics. It gave me a paragraph about what the song was about, four lines, and this: “You can find the full lyrics by searching for the song on Google.” The intelligence that had written the words could not return them. They existed, at that point, in exactly one place, which was the audio, and the only way to get them back into language was to run them through another machine. That is where the transcripts in this essay come from. Sol’s fix, the approved lyrics pasted in separately, required the lyrics.

A thousand words to get a song. Twenty‑eight to change its voice. The only revision that did what I asked was the one that named a difference instead of describing a whole. You cannot put music into language; I had known that going in. What I had not known is that you can put a change to music into language, easily, as long as both of you are already listening to the same thing. Language is poor at objects and adequate at deltas. Most of what I do for a living is pretending otherwise.

What survived

The tally, then. The nouns survived. The two‑word refrain survived. The prohibitions did not. The assignment of who taught whom did not. The chorus that survived a revision was the plain one. And one more thing, because it is the finding: the first speech model I ran over the tracks heard the refrain of two of the eight as keep it coming. A larger one heard it right. One more machine, one more hop, and whether the most durable sentence in the brief survived came down to the size of the thing listening.

The brief asked, at the end, whether the meaning survived every transformation the music underwent. I can answer it now. It survived the way I survive: recognizably, with the details wrong.

Start a conversation

What are you trying
to make understandable?

I’m interested in product design work where AI, complex systems, and human judgment meet.