I played around with making a bigger system, and asked Thaura for advice. It seemed to work faster, and produce cleaner results, sometimes.

I was trying to figure out how to write larger programs, specifically, how to develop parts independently, and then combine them. Normally, you use libraries, but with this LLM coding, I felt like I’d be in and out of modules to improve them, so I needed something bigger. I asked for advice:

https://thaura.ai/?share=f4c0b593-4b5b-4594-957c-5ae83d030e97

It read some OpenCode thing, and then proposed a structure.

I tried to make a test app and learned a few things:

• You can tell it to look at prior projects, to copy features. This is really good, almost magic.
• It created an inbox/processing/done folder structure for me, and I tried to break the processing steps along those lines. A prior project did this, too.
• It likes the orchestrator pattern. It likes the strategy pattern.
• It likes sidecar files.
• DRY: code first, and then study it to find repetitious parts, and generalize that code into a library.
• Ask the model what packages to install, and install them. If you don’t install them, it’ll try to write it’s own version, or try and manage a library folder.

During this session, I didn’t put in real world data. So the code isn’t going to work well for real world data.

I also didn’t have a real world use case for what it built. So it’s kind of a toy program. Still, it uses NLTK, Tesseract, and some other libraries I don’t know, so, I’ll learn some of that.

Here are the vibe coded files. I haven’t read them, but that’s for tomorrow.