HP – What did you do, this Summer?
AE – Created a music plugin with Claude.
HP – Didn't realize you were a coder! Who’s Claude?
AE – I'm not. Claude is an Artificial Intelligence system from Anthropic.
HP – OIC! You’re among those people who anthropomorphize AI…
AE – Nope! The same way you don’t anthropomorphize when you say “I took a walk with my dog” or “she smashed this rock with a sledgehammer”.
HP – Fair. So, you made a plugin, eh? Care to show me it?
AE – Here ya go!
HP – Wait! Isn't the circular sequencer a Native Instruments thing?
AE – It’s a Hy-Plugins thing, and Olympia Noise Co. thing, and “Buchla USA” thing… The circular representation is just a convenient way to display cycles… and it’s in use since at least 1252.
HP – Right, so, what makes your plugin special, then?
AE – The fact that it came from me? I mean, it doesn’t need to be unique. It’s a project, not a product. In fact, I’ve put it in the Public Domain (CC0). And it has plenty of features that I find interesting, for both “didactic” and performance purposes. So, it’ll be part of my playflow.
HP – You mentioned it’s free…
AE – As in speech.
HP – Can I try it?
AE – Totally! Well, if you have a Mac, totally. You can download it here from the release page and install it directly. For iPadOS, I hope to have it on TestFlight soon.
HP – Neat! Still… Nerdy way to spend a Summer, no?
AE – You’re not wrong. I’m a learning geek. 

The Plug

HP, Hypothetical Person, uses a Mac and has access to plugin hosts and DAWs, including the “free as in first-party perk” GarageBand. The core work behind that plugin is available as a webapp.

What do these do? Simply put: you enter a pattern and you get it displayed in a circle. You can also play it, to hear what it sounds like.

If you don’t know anything about music, you may find this confusing. If you’re still with me, great! I could use your fresh perspective.

If you know a few things about music, you may find this confusing. At this point in the project, I haven’t done much to make things immediately obvious.

If you know a lot about music, you may find this frustratingly confusing. I get that.

Many of the things I choose to do in this work are rather unusual. And they come from multiple origins. The plugin and webapp both have disparate use cases.

Getting the vibe

For a while, now, I’ve been playing with computer code using Large Language Models (LLMs). There are many approaches to generating computer code, some of which qualify as vibecoding. And there are many ways to understand vibecoding. As with the old joke (“ask ten people, get 15 different opinions”), we could argue to no end as to what “it actually means”.

To me,

Vibecoding is a practice of creating fully working computer software through interactions with a Large Language Model without getting into the details of the code itself.

So it’s not really “coding” in the strict sense. Sure, it does involve a number of steps which are quite similar to those involved in software development (which is the whole thing which surrounds coding). Debugging/troubleshooting, especially. Yet it’s quite different from coding tasks in software development projects. (Yeah, I know that coders spend a lot of time at these other tasks. They’re “part of” coding, in that sense. What’s missing in vibecoding, though, it’s the part which involves, y’know, getting code together.)

Otherwise, it’s much closer to “design” in the sense that we give specifications as to how things should behave and interact based on those results. It’s also akin to management, though it’s remarkably different in key ways.

Basic idea is: “it’s a vibe”.

While the practice has been going on for a while (including by yours truly), Andrej Karpathy coined “vibe coding” this past Winter.

There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like “decrease the padding on the sidebar by half” because I’m too lazy to find it. I “Accept All” always, I don’t read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, l’d have to really read through it for a while. Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away. It’s not too bad for throwaway weekend projects, but still quite amusing. I’m building a project or webapp, but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

Being at home with Claude

(With a nod to a cousin for a leading role.)

So far, this project has been almost entirely done through interactions with Claude.

It started as a simple artefact through a chat with the Claude AI chatbot. It grew more complicated as time went on. Eventually made it into an open, trackable project as a repo on GitHub, the Microsoft-owned platform used to share many an Open Source project.

Since then, mostly been using Claude Code, the Anthropic tool for computer code. Though it’s really geeky, I find that it suits my workflow.

It’s been an intense learning experience. Not in a bad way. In a true way. 

I’m kinda proud. Feeling accomplished. And, yeah, learnt a lot!

As with a lot of deep (human) learning, much of it is difficult to describe. Some of it is demonstrable. Bits can be reproduced. 

Hard things are easy, easy things are near-impossible

Of course, there are difficult things which remain difficult. It’s just that… you think something will take a huge amount of time and it’s done in an instant. Then, when you want to make a small tweak… you’re in it for quite a ride.

Take the initial Claude artefact. Within seconds, I had something working. The weeks since have been a constant refinement, of course. Yet many of the features which took the longest were simple tweaks. Were I a coder, I’d likely be able to solve them quickly. Alas, I am not.

The strict left-to-right order I mentioned in the nerd block? Fiendish. Kept causing errors. Claude would not cooperate. It was just about flipping things around, an easy task in computing. Yet it proved nightmarish and haunted me for weeks. (Had to do it over again, at some point in the plugin development. And, as I type this, I still have a “todo” item to change the original webapp implementation.)

Recently, I had to spend hours fighting Claude with an issue caused by ‘E’ being interpreted as ‘e’. Claude Code turned the whole house upside down, making things fail miserably, only to find the lowercase issue.

Life is a test

Learnt a lot about the importance of testing. Had heard a whole lot about “unit tests”. Eventually implemented a series of them. They’re far from sufficient, especially when it comes to user-facing features. Still, they’ve proven quite useful as a way to eliminate some recurring issues. In fact, it’s funny how Claude Code will react while creating a unit test which clearly demonstrates a flaw in the code.

Testing is at the very core of LLMs and GenAI in general. At a very basic level, a significant part of the training is about exceedingly large numbers of tests, with clear results.

And that’s the key. If something is easy to assess automatically, it’s likely easy to generate. Humongous numbers of iterations can get you a long way, when you can get clear feedback.

On the other hand, those things which are difficult to assess (the majority of Life, I’d say) is exceedingly difficult to get through most generative tools. You could possibly make a simulacra of the thing through an automatic assessment along some robust criteria. And that’s how we get things which give the impression of being created by humans. Yet it all lacks intersubjectivity.

Know thy self

As is so frequently the case in learning that it sounds trite, my vibecoding process is also about self-learning. That gets tricky to explain because it’s about reflexivity. Still, it’s… telling.

Because I’m creating my own plugins, I notice that I don’t feel the same way about commercial products in the whole #MusicTech scene. In there, as in other hobbies, there’s a notion of “GAS”, Gear Acquisition Syndrome. With frequent jokes about “a cure for GAS”. While I’ve bought a few things over the Summer, my attraction to most of these products has really shifted. Which even gets me to dream in new ways, both figuratively and literally.

Not only am I not a coder, I’m quite vocal about not being a manager. Sure, I can (and often do) lead projects. That’s often called “project management”. I cannot, however, manage people. It’s even about psychological demands. Thing is, some part of interacting with Claude feels to me like management. Not that Claude Code is “like an intern”. Certainly, LLMs aren’t like learners since they simply do not learn. I don’t mean that they’re reluctant learners. They don’t (currently) learn from our interactions, though some of them may refer to those interactions when instructed to do so. (Huge difference with human learning.)

Also learnt some hard lessons about knowing when to “cut my losses” and start over. Been reading Amy Edmonson’s Right Kind of Wrong. Quite fitting. When Claude Code gets “carried away” (its words!), it’s about holding back. Here, as in many other aspects of the process, versioning through Git has been key. Can be tricky and I’m no expert at that. Still, I’m already way better at it than I was a few weeks ago. It counts as self-knowledge for a variety of reasons.

Vibing to the beat

Also been learning stuff about music through this project. Some is very obvious, such as understanding some geeky algos by getting them implemented in my software. Or getting inspired from a math/music video.

Then, there are things which happen unconsciously. Such as hearing beats in their representation. As I’ve been listening to contemporary Pop music while doing cardio training, there’s quite a bit that I notice in a new way. An example could be the so-called “Scotch snap”. In my notation, it’s 0x3. And, thanks in part to the Baader-Meinhof Effect, I get to notice it more than I would otherwise. Same with the Tresillo rhythm (0x94) which is often used in the documentation which has been inspiring me through much of this vibecoding process.

Fieldwork vibes

Which is a good transition into the connections between this whole project and my background in ethnomusicology. And that one is a bit weird.

Some responses from Claude have alluded to a very direct relationship. After all, documenting musical patterns from diverse origins does connect with some aspects of ethnomusicological work (though it’d be much closer to comparative musicology than to my approach/era in ethnomusicology). In fact, there’s a kind of tension involved, as there’s a profound transformation at stake when we notate music.

As anyone can sense, I’m quite critical of such work as an ethnomusicologist.

What I like about those things, though, is that they can be fun bases for experiments. In fact, it’s satisfying to play with them. The rhythms they produce range from boring to funky to confusing to puzzling. Using algos to create patterns is as good as any a way to play with music.

So, yeah, I’ve implemented several features from those algos. Including “opposite” versions which give results which are uneven and/or imbalanced and/or dispensable. Plus, tweaking rhythms using such tools works well in my playflows.

At the end of the day…

This is all part of “my idea of fun”.