Having spoken a few times about my experience with AI ‘vibe coding’ (as I’m told it’s being called…) I thought I would drop a post with my latest experience, being that it’s good news for anyone else wanting to give it a try.
I can imagine experienced ‘today’ coders laughing at my innocence in all this, and that’s fine. I come from a world of BASIC, assembler and COBOL. OO languages, despite my several attempts to learn, still absolutely baffle me. So too does the modern development cycle. So let me be clear that I understand neither the code I’m using, nor how the development process works, yet…
The bottom line is that I have found a way of coding up a native VST3 plug-in that has required no local installation of IDEs, Juice, or compilers on my own computer. The whole thing takes place in Github, for the sake a few copies and pastes.
For reasons I needn’t bore anyone with I wanted a plug-in that takes the MIDI output of an MPE controller and converts it to data suitable for a non-MPE synth with poly aftertouch.
After a fair bit of prompting, re-prompting, and by combining the efforts of both Gemini and DeepSeek, I was given the keys to the door of Github and four chunks of code. Once the code is ‘committed’ in Github the compilation kicks off automatically and out the other end (unless it finds an error) drops a VST3 plug-in, ready to go. Alternatively I give any compilation errors back to my AI friend to try again - and back around the loop we go.
Now I have the plug-in I needed and the basics to do the same again any time in future. I don’t know how it works, but I don’t need to. As long as I can explain what I need I stand a fighting chance of success.
I’m not going to explain the process - AI can help with that, after all! It can talk you through the steps much better than I can. All I’m here to say is that it can be done. Maybe even give it this text for a starter.
Thanks for the info. I’ll look into this as I am currently vibe-coding two plugins, one is a MIDI input transformer to trigger a sequence of chords one step at a time from a single button, the other is a synth for bass guitar that can track a fretkess 5-string. I’ve been using Chat GPT and Juce and both plugins are working surprisingly well.
I don’t (yet) have a dog in this hunt, but just a couple thoughts. First, many of us view music in general as a hobby, maybe a small scale business. I think having support that can program things makes good sense. It’s a niche business (how many musicians who want software solutions are there?), many want to get involved for fun or profit, but may not understand the wants and needs of the musicians, and can make the changes needed as quickly as machine coding. How many can understand both the musical side, and the code side? I have considerable PC experience, but mostly in other areas, l cannot code on any system, at this point. Maybe after l retire.
I guess my main point would be, for the hobbyists among us (many), anything that helps make our hobby more affordable, flexible, and easier to pursue is a good thing. I am no kind of audio purist, and l am not going to gate-keep for much more expensive gear, so make of that what you will.
Nice. I need to try that, I want a VST that basically works like the velocity mapper in Pianoteq, ie one that can easily be calibrated and where I can save different setups and use as profiles.
It’s amazing, isn’t it? You work your way through a song in a DAW or VST host like Cantabile, and find yourself wishing for a simple way to accomplish one task. Whereas before you would dig through tutorials, manuals and help files and maybe find a way to do it in a round about way, now you can ask for the tool and have it made for you almost instantly and at no charge in many instances.
I had a VST3 plugin made by AI just for my own use and plan on getting more of them done in the same way.
It’s certainly life changing for the software user and the professional software developer.
I would never dream of attempting to vibe code something really deep and complex. For that I just look on the market. It’s faster and easier to just pay for complex software. But little things that make one’s life easier? Absolutely.
It’s these little pieces of code that can scratch those itches many of us have.
A few years ago a kind chap here coded me a plug-in that would quantise key presses to a bar boundary. That’s the kind of plug-in that nobody is going to create for fame or fortune, but it’s one of my most used. I’m probably the only person on the planet using it.
Now I’m self-sufficient in itch-scratching plug-ins.
I think vibe coding is great for small things like this. Big companies seem to backing away from AI programming now, as they have realised not knowing how the code works is a maintenance nightmare and potentially leads to maintenance, security and data access risks
AI is improving very quickly though. I used it about a year ago to script some functionality in Q-Sys and it was not a very smooth experience. The first version of the script worked, but then asking for changes quickly deteriorated the script to the point where it became unusable. Then I tried AI to assist me with a 3D printing project and the same story happened. Almost a good result at first, then edits breaking it instead of improving it.
In the last two months I tried it again with a 3D printing task and it went extremely smooth. Without prompting for it, Claude was smart enough to include a few variables in the code and explaining them to me so that I could just type in a couple of values and change the look of the part I was making. It also suggested rounded corners of just the right size without me prompting for them. It was like working with an experienced pro as opposed to a robot that follows instructions mindlessly.
Then I had the VST 8x8 mixer for my Cantabile setup vibe coded and it worked on the first try. The few edits I asked for did not break it.
I definitely see the improvement in the overall performance of AI that was reported to have happened in the last 6 months.
The initial hype built around AI has hurt it because it could not deliver for a long time what was promised. It seems to be catching up to the hype recently.
It is indeed improving all the time, but I have seen examples where “vibe coding” has provided code with huge gaping security back doors, and in breach of GDPR regs, which would be seriously bad news in my line of work, and as, I’ve said, some companies have come to realise it comes with a hidden maintenance burden cost when users start reporting bugs.
So, it is very much “horses for courses”, and that’s not to say it will not get better over time of course.
I must confess I am an AI fence sitter. I am using it slowly in the right context that suits me. I’ll let the “rest of the world” be “bleeding edge”
Of course if it works great in your context, like Brad using it for test script automation, or the examples given above, that is all to the good.
My point is: make use of it, but do it with your eyes wide open about the fact it it can lead to issues such as I have described, and consider if that matters to you or not.
I support your cautionary note, but I don’t think those kind of dangers are going to be reflected in a couple of MIDI plug-ins for my own use. Anyone thinking beyond that boundary… buyer beware.
I’ve been a professional software developer for over 25 years, and the last 6 - in fact, the last 3 - months has seen me totally change my workflow. Whereas before I’d spend a lot of my time looking up options for how to do something (even if the code itself wasn’t that complex to write) now I get AI - mostly Claude Code - to do the heavy lifting. I’m now more like a manager / consultant / quality assurance person for the AI.
Of course, I have the advantage that I can look at the output and critique it with (hopefully!) a bit of knowledge. I also know how to ask the AI to do things in a way likely to get good results first time, as I pretty much know how I want things to be written.
I agree with you. The danger is when AI gets it wrong and people fail to spot it because they no longer have the experience to spot it.
My line of work involves hazardous activities management, where the consequences of error could be fatal to both workers and the general public.
Our AI experimentation outcomes so far can be summarised as: promising, but sometimes erroneous with potentially fatal consequences if that was not spotted by the “meatware”. That is all well and good when there are (allegedly) wizened old gits like me about…
Don’t get me wrong, I think AI has great promise, but for me it is a cautious approach of finding where you can use it with confidence and where you need to be very careful…
