I Copied My Own AI Music Specification. It Couldn’t Reproduce My Work.

A practical FZ-SA-01 experiment showing the difference between a specification, a generative tool and a developed creative process

There is one sentence that appears constantly in discussions about AI-assisted music: “You just write a prompt and press Generate.” Fine. Let’s test that instead of arguing about it. Not theoretically, not through another debate about who is or is not a “real artist”, but through a very simple practical experiment: I took my own previously developed sound architecture specification and transferred it into a fresh generative environment. If my entire creative contribution really exists inside a text instruction, reproducing the result should be relatively simple. It wasn’t, and that is where the conversation about generative music becomes far more interesting.

This is not simply a prompt. It is a specification.

Within FocusZen Code, I stopped describing what I work with as simply a “prompt”. I developed sound architecture specifications, including FZ-SA-01 and FZ-SA-02, which define the structure, direction, behaviour and character of the sound I am trying to build. That distinction matters. A basic prompt might say: “Create a dark electronic track with deep bass and a female vocal.” A specification is something different; it is a design for what I am trying to achieve, describing relationships between elements, structure, instrumentation, spatial behaviour, dynamics, vocal characteristics, constraints and other components that together create a particular sonic architecture. I do not publish my complete specifications because they are part of a creative workflow I have developed through months of work, and I see no reason to publish the entire recipe simply to prove which tools I use. What I can do, however, is something much more interesting: I can test what happens when I copy my own recipe.

The experiment: take the existing work and simply press Generate.

I took part of the FZ-SA-01 specification that I had already developed and used during my previous work. I did not create a new instruction specifically for this experiment, and I did not try to describe the desired result again using different words; I used material I had already developed. I then transferred that specification into another AI music generator and pressed Generate. Music was generated, but it was not my music. I then performed another test using the same principle in a fresh Suno environment, without the established workflow I had developed through my normal work. Again: specification → generator → Generate. Again, the resulting output did not reproduce what I had previously created through FZ-SA-01. Put very simply, copying and pasting an already developed instruction was not enough.

If the prompt were the entire work, this experiment should have ended differently.

This is where one of the most common arguments about AI-assisted music begins to fall apart. If my work consisted exclusively of finding the right words and writing the right prompt, we should be able to COPY → PASTE → GENERATE and obtain essentially the same work, because according to that argument we have transferred the element containing the entire creative value of the process. Instead, we get something else. The generator executes an instruction, interprets it and produces an output, but an output is not automatically the intended creative result. Anyone who has actually spent substantial time working with generative music knows what comes next: no, that is not it; try again; the vocal is wrong; the bass behaves incorrectly; the structure collapsed here; that transition does not work; this version contains one useful element but everything else needs to go; again; change it; test it; reject it; go back; compare it; adjust it. Suddenly, what was described from the outside as “she typed a prompt and pressed a button” becomes dozens or hundreds of individual decisions.

This is the part of the work nobody sees.

The listener sees the finished track, but they do not see all the versions that were rejected, they do not hear the generations that completely failed to match the intended design, they do not see the changes in instructions, the testing of relationships between different elements or the moments when one apparently small change destroyed the entire structure, and they certainly do not see the months spent learning how the tool behaves. I also want to be precise about what this experiment demonstrates: I am not claiming that this test technically proves that Suno changes its internal model weights specifically for me or individually “learns me” in the machine-learning sense, because proving that would require information about the internal architecture and personalisation mechanisms of the system. What I have demonstrated is something simpler and directly observable: my specification, without my developed process around it, did not reproduce my finished work.

Months of work change the way you use the tool.

In everyday language people sometimes say that they “train the AI”, but technically it is more accurate to say that we train our own ability to work with a generative system and develop a workflow between human and tool. That is what I have been doing for months: learning what particular instructions cause, how the generator interprets certain combinations, which elements work together, which do not, how far a particular component can be pushed, when the direction needs to change and when a result needs to be rejected even if it sounds technically good because it does not correspond to what I intended to build. It is a strange kind of instrument because you do not press C and reliably receive C; you provide the instrument with a set of instructions and receive an interpretation, which means that you have to learn not only how to give instructions but also how to anticipate the behaviour of the system, recognise useful outputs and guide successive iterations towards your own creative vision. That is a skill, and it does not arrive automatically with the Generate button.

A specification is not a finished track.

A more accurate representation of the process looks like this: CONCEPT → SPECIFICATION → GENERATOR → INTERPRETATION → LISTENING → HUMAN DECISION → REJECTION / CHANGE → NEXT ITERATION → SELECTION → FURTHER PRODUCTION → FINAL RESULT. The generator exists inside this chain, but it is not the entire chain, and that distinction matters. FZ-SA-01 defines the architecture I am trying to achieve, the generator attempts to interpret that architecture, I evaluate whether the interpretation corresponds to the project, and if it does not, I continue working. Possessing my specification therefore does not automatically mean possessing my process.

We already understand this principle in traditional music production.

A producer can publicly say which DAW they use, show their studio and explain which synthesisers or plugins they work with, but that does not mean they are expected to give everyone the complete Ableton project, every preset, every compressor setting, every automation curve, every effects chain and the exact sequence of operations used to create the final sound, because part of their craft exists inside that process. Generative AI production is beginning to reveal the same distinction. I can say that I use generative AI, I can show the finished result and I can even demonstrate an experiment using part of my specification, but that does not mean I am required to publish the complete mechanism I spent months developing. Transparency about the use of a tool and publicly giving away your creative know-how are two completely different things.

“AI made the song” is therefore an incomplete description.

This is not about pretending that a generator does not perform an enormous amount of computational work, because obviously it does, and it is not about claiming that everyone who opens an AI music generator automatically becomes an artist, because they do not. But the opposite conclusion does not logically follow either: using a generator does not automatically mean that a human stops being a creator. The more useful question is not simply “Was AI used?” but “What decisions did the human make between the original idea and the final result?” Two people can receive exactly the same tool, and we can even give them the same specification, yet they do not necessarily create the same thing. My experiment demonstrates precisely that distinction.

I copied my own instruction. I did not copy my own work.

I took part of my own previously developed sound architecture, transferred it into another environment and pressed Generate. The generator generated music, but it did not generate what I had previously created. Something was missing between the specification and the finished result, and that missing element was the process: months of experimentation, selection, rejection, adjustment, listening, decision-making and developing my own way of communicating with the tool and understanding its responses. So the next time someone says, “AI music is just a prompt and Generate”, perhaps we do not need another argument about it at all. Copy the specification. Press Generate. Listen to what happens. Pressing the button is easy. Getting the tool to produce a result that matches a creative vision that previously existed only in your head is the work.