AI Coding for Pure Data

Organelle Patching
1

Hi Everyone,

This morning I tried to convert my MIDI Snow Generative sequencer that I’ve created in HTML/JS MIDI Snow — Sonorlab into a PD patch using Claude Sonnet 3.7.

To my surprise, it seems that it can actually write PD Text Code but it only goes about 80% because it keeps missing connections between objects in PD. No matter how much you try to tell it to fix its mistakes by giving it screenshots of the patch, it misses a lot of stuff every time it writes new code. Since this is quite a complex patch to create (at least for me a total PD noob), I was hoping to get it done but now I’m stuck with an almost-working Organelle patch. If someone knows a good prompt to avoid this type of almost-complete code creation or wants to collab on this sequencer / help me finish it let me know!

2

I have a feeling that this is a symptom of LLMs working in words not pictures. One thing that might be worth trying is to get Claude to write the patch in python using this: GitHub - dylanburati/puredata-compiler: A tool for writing PureData patches then “compile” to pd yourself afterwards.

Good luck!

3

Indeed, when I tried this a year or so ago, AI kinda got the objects and dreams up the connections.
An IT specialist told me stuff might improve as the AI training just started. But I’m happy I gave up as the programming work itself became most of the fun and I wouldn’t give that part of the creative process away.

2 Likes