Creative

How Jeffrey Emanuel Uses Opus 5.5 to Arrange Virtuoso Piano Music

Built by Jeffrey Emanuel · · Source: X

Quick answer

Jeffrey Emanuel built mtdt, an AI music theory tool for agents, plus a library of agent skills. One skill takes a simple melody and turns it into an over-the-top virtuosic piano arrangement. They tested it with Opus 5.5 on Hagood Hardy's "Anne's Theme."

Result: The author says Opus 5.5 "did a yeoman's job" with the task and produced a showy "Ultra-Schmaltzy" arrangement that is meant to be playable by a human. They still think the original sounds better.

What they built

Jeffrey Emanuel built a music theory system for AI agents. It has two parts: a software tool called mtdt and a library of agent skills that goes with it. One new skill takes a plain version of a melody or song and makes it far more virtuosic. Their model for this is Vladimir Horowitz's Carmen Variations.

How it works

Emanuel says most of their earlier experiments ran on GPT-6 Astra. They tried Opus 5.5 because it costs much less. The input was "Anne's Theme" by Hagood Hardy, from the 1985 Anne of Green Gables movie. The system turned both the original and Opus 5.5's version into animated score videos. According to Emanuel, the system checks every piano piece carefully to make sure a real human can play it. The post doesn't explain how mtdt or the skill is set up.

The result

Emanuel says Opus 5.5 is "really good at musical tasks" and did "a yeoman's job" given the brief. The output was an "Ultra-Schmaltzy" Horowitz-meets-Liberace arrangement that they think only a pianist like Lang Lang or Kissin could play well. They still prefer the original musically. They describe the exercise as a test of how far the system could push a piece while keeping its essential character.

How to build something similar

  • Give the agent a structured way to read and write music, such as a notation format or a music library, so it isn't working with loose text.
  • Write each transformation as a reusable skill, with a clear brief and a reference piece that shows the style you want.
  • Add automatic checks for your constraints, such as whether a human can actually play the piece, and run them on every output.
  • Run the same task on more than one model to compare quality and cost.

The original post

OK, I can officially confirm that Opus 5.5 is really good at musical tasks! So far, most of my experiments with my new AI music theory system (a software tool for agents called mtdt and its corresponding agent skill library) have used GPT-6 Astra, but I was eager to try Opus 5.5 yesterday because of the much lower cost. As a sample task, I had it use a new skill I created. The basic idea of this skill is to take as input a simple, straightforward version of a melody or song and then make it way more virtuosic in a kind of over-the-top way. My prototype and exemplar for doing this well is Vladimir Horowitz's version of Bizet's Carmen (Google "Horowitz Carmen Variations" and you can see the video; it's pretty mind-blowing if you've never seen it before!). Now, I'm not going to argue that this process makes the music "better," but it's certainly fun and a great exercise that can demonstrate a wide variety of music theory fluency and compositional methods, techniques, and devices. I gave it as input the lovely "Anne's Theme" by Hagood Hardy from the 1985 Anne of Green Gables movie (see my post yesterday about this movie). I've used my system to prepare two animated score videos, which you can see below. The first is the "plain" original version, so you can first understand what the system is starting from, and the second is Opus 5.5's attempt at making a virtuosic "Ultra-Schmaltzy" Horowitz-cum-Liberace version that you'd probably need to be Lang Lang or Kissin to play well! The last image shows part of the score for each one side by side, with the original on the left. So what do you think? I'd say it did a yeoman's job here given the mandate. I'll post a link to the PDF of the score if you want to try your hand at playing this beast yourself (if you can do it, I'll be impressed! Note that it is designed to be playable by a real human; my system checks that carefully for any piano piece!). For the record, I think the original is better musically. This was more of a test to see how crazy the system could make things while still preserving the essential character of the piece.
Jeffrey Emanuel (@doodlestein) · September 23, 2026 · View on X

Want to build something like this?

Join Shipping Skool free: 4 live build sessions a week and a community that ships.

Join Shipping Skool — Free

More use cases

← All AI agent use cases