Scrutable

I expect this may be a divisive topic, but...

I did my first project with Claude Code at the weekend. I haven't coded since my 2nd year at uni using Fortran 90. I've read about what Claude Code can do, and have been encouraging my team's developers to experiment with agentic coding. It only seemed right that I spend some time actually playing with it to see how it works.

It was easier than I thought. As well as Claude Code itself I used the main Claude chat to walk me through the initial setup of Visual Studio, GitHub, Python etc. It gave me the prompts I needed to then get Claude Code running itself.

The project I chose was a music stemming solution which splits an MP3, WAV or FLAC file into individual instrument tracks, including one combined drumless track which I can use when doing my drum covers. The actual stemming algorithms are open source (demucs, built by Meta).

First step (after initial setup) was getting it to build something that would take a file, split it, and output to a folder using a command prompt. It also needed to work with a YouTube URL, scraping the audio and splitting it.

When I was comfortable that worked I then go it to build a UI for it that includes a library of previously stemmed songs and a playback feature that can toggle between drums only or drumless.

It's not the prettiest, but it all works. Until now I've been paying a subscription for a stemming tool which I can cancel. This also does a couple of extra things I wanted:
- works with a YouTube URL
- creates one single drumless mix

The final step was to connect Claude Code to GitHub so it could push the project to there and create a .exe version for PC users.

At no point did I need to write any code (I can't). Any errors were fed back to it for it to interpret, find the issue, and update the code. This included when the .exe build failed in GitHub. Claude Code was able to see this, identify a missing dependency, and fix it before pushing the project back to GitHub to try again.

I want to experiment with it more and am trying to think of useful tools I could build. I have a slightly more ambitious one in mind next. Using Claude to help scope it out based on an initial broad brief has been helpful.

Has anyone else played around with it?

Here's TriStem:

Cool tool...

I'm my industry (simulation) there is a guy posting regularly on this use of Claude to connect various disparate open source simulation tools. He's generating some good results with them. He's trying to attract venture capital interest.

There are also other AI startup's coming to my industry.

Where it leads... who can say. It's one thing for AI to empower hobbyists, it's another for commercial companies using AI alongside commercial tools operating on commercially sensitive data.

It seems to me that the central blockage for commercial applications is that in order to train an AI agent requires data, and commercial companies don't want to share their data. The data issue is reminiscent of when cloud storage first became available, but whereas you can ringfence cloud storage and it still be useful, I would imagine erecting barriers between agents to protect customers data would make them a lot less useful, at least in our niche. Although I could imagine some basic agent was provided and then the customer customizes it with their own cases and runs it locally to them - but only big customers would likely have the desire to do that, and there may not be enough of them in our industry to make it viable.

Anyway, my company is still very much trying to figure out where AI makes sense... and we're far from a conclusion. Our customers aren't requesting any AI features as yet.

Leaving aside the wider issues of how generative AI models are trained and how they are being applied, that looks like a nifty tool. A quick web search suggests to me that there are existing free packages that would do the same job, but this one is customised to what you want and you (kind of) built it yourself, which is always a nice thing to know.

Once built does it just run locally on your PC, without depending on any online services? I might be slightly paranoid but I would be very cautious about running any software created in this way on a computer that also handles any personal (or work) data, including your email. I've not heard of AI code tools building actual malware in to the code that they write, but I have heard of them using malicious third party libraries that were designed to trick AI code tools into using them, and every so often legit third party libraries get poisoned with malware, so if you haven't manually checked which versions of everything your tool is using and checked them against a vulnerability database then you don't really know that you aren't at risk. If it doesn't need online access you could run it in a virtual machine configured with no networking, for example, but that's introducing more techy stuff counter to the idea of not needing to know much techy stuff, obviously. Until I'd researched things a lot more thoroughly I would not use an AI tool to build anything for deploying to a cloud platform like AWS, because securing those things properly is not a job for beginners - I work on such products and know enough to know that I don't know enough to do it. In case anyone doesn't know, do not under any circumstances use these things at work except as explicitly approved by your organisation - you are sending data to a third party who may well have granted themselves the right to use it for their own purposes unless you have a contract with them not to. I've read about people in other companies finding people have invited AI 'meeting assistant' bots to their work calls off their own bat to 'assist' them, which is just (head-exploding-emoji).

We are being encouraged to do stuff with these tools at work, and our customers do think they want AI features in our products although I'm not sure if they really need them, so it's something I need to get to grips with this year, maybe starting with something low-stakes like test case generation.

I think the most immediate issue with these tools is that their business model doesn't seem to be sustainable, because it costs far more to provide them than they are making off them in revenue and sooner or later they are going to run out of investors' money. What happens at that point I dunno.

nekomatic wrote: Mon Feb 23, 2026 8:38 pm Once built does it just run locally on your PC, without depending on any online services?

First time it runs it does need to download the demucs stem splitting algorithms. But after that it's all being run locally.

We are being encouraged to do stuff with these tools at work, and our customers do think they want AI features in our products although I'm not sure if they really need them, so it's something I need to get to grips with this year, maybe starting with something low-stakes like test case generation.

I see this as two separate things:

1) using these tools to build stuff, either professionally or personally
2) building AI tools themselves

At work we're doing both. We're using CoPilot and other AI tools extensively internally and these have been really useful. Our devs (mainly our products teams) are using code assistants like Claude Code, and I'm encouraging mine (professional services devs) to do so too.

But we're also releasing AI tools to our customers too. We don't really have an option not to in our space, and many of our customers are interested and enthusiastic about it.

I don’t touch the things unless forced to, I think they are deeply ethically tainted. FWIW I’ve studied machine learning and AI at Imperial College so it’s not out of ignorance.