I can't compose, yet I made a piano piece that plays itself: what's truly scarce is never skill, but judgment

*▲ I give the direction and AI implements it. What held me back was never musical ability, but my belief that I needed musical ability. *

What is truly scarce is never skills

I can’t compose music. I can’t read music notation, I haven’t written a single line of Web Audio, and my understanding of sound synthesis is at the level of “the speaker will vibrate.”

But a few days ago, I made a web page that plays original piano music by itself. The falling notes, the lit keys, the speed can be adjusted, and can be exported to a sound file for storage. They are all in one file. Now it is hung on [a URL] (https://piano-falling.pages.dev/), and you can open it directly to listen.

In the past I would have said it was “something I can’t do.” Now what I want to say is another thing: what blocked me was never musical ability, but that I thought I needed musical ability.

This difference is my deepest experience with AI in the past two years.

My role has changed from “executor” to “judgement person”

Many people’s imagination of AI creation is: my next sentence will produce a finished product. But when you actually do it, you’ll find that’s not the case at all.

When it came to composing the song, the AI ​​didn’t decide anything important for me. It asked me: What key do I want? What emotion? How to arrange the form of music? I had to first translate the vague idea in my mind, “I want a piece that has ups and downs and will be pulled back at the end,” into a language that it can perform: A minor, a three-section structure (the verse is laid out, the middle section advances, and the ending closes), the chord progression Am–F–C–G is a classic progression, and a borrowed dominant chord is used in the middle to create tension, and then the melody is pushed to the highest note as the climax.

You see, none of these judgments were made by AI for me. It is responsible for turning my judgment into executable notes; I am responsible for knowing “what sounds good” and “where to be tight and where to be loose.”

This is exactly the point: when the cost of execution approaches zero, all that’s valuable is judgment.

Just because you don’t understand technology doesn’t mean you can’t describe what you want.

Sound synthesis is an area completely foreign to me. But the interesting thing is that being unfamiliar has not turned me into a Party A who can only say, “Help me make it sound better.”

I can tell what I want: the treble of the piano should be grainy, the bass should be thick but not muddy, and the harmony pad should be soft underneath. AI translates these descriptions into technology: the triangle wave is stacked with an octave overtone, the bass is passed through a low-pass filter, ADSR is used to control the rise and fall of each sound, and a layer of artificial reverb is added.

What I learned was not the instructions for Web Audio. What I learned is this: When you can describe exactly what you want, you’ve already done half the job. Technical details can be outsourced, but taste cannot.

Furthermore, this also explains why when using AI, some people make very rough things while others make very fine things. The difference is not in the tools, but in how accurate the ruler is in your mind.

Boundaries are for pushing, not for recognizing.

The most interesting thing about this little thing is that it keeps growing.

At first it was just a moving toy. I looked at it and thought: Can I adjust the speed? Can you just click the keys and play randomly? The next question is: Can I record the music and export it into a file for storage? Can you put it online for others to see?

Every “can or not” is a boundary that I originally thought. But every time the answer is “yes”. To export audio files, I thought I needed to learn audio engineering, but the result was that the browser was able to render the entire song offline into a lossless file. When I went online, I thought I needed to rent a server, but the result was a command and a free URL.

The more I do it, the more I understand: most of what we think “I don’t know” is actually “I haven’t tried it.” There is now only one question between the two. And pushing an idea all the way to the line is what is really fascinating about Vibe Coding.

For you who also “want to do it but don’t know how”

If you also have a list of things you’ve always wanted to do but got stuck on, let me give you a specific suggestion.

Pick the smallest, least important, and most play-like one above and make it with AI today. Not for the work itself, but for something more important: recalibrating your understanding of “what I can do.”

Because for One-Man Company and for any creator who wants to use AI to amplify themselves, the most expensive cost is never the tool fee, but the line in your heart that “I can’t do this.” Wherever you draw that line, your world becomes bigger.

I can’t compose music. But now I have a song of my own, I can play it myself, and I will keep playing it.

That’s enough.

---

If this piece stirred up a “something I’ve wanted to make but can’t” in you, don’t let it stay just an idea.

I’ve packed my full workflow for taking an idea from zero to live with AI into a course: the Vibe Coding for Claude Code hands-on workshop. In three hours, you’ll use the CLI to turn an idea into a real, usable, shippable digital asset with your own hands. If you want the ability to push any idea all the way to live, come build alongside me.

---