Member Avatar for jkon

The company I work for uses SVN, with full control over backups on other servers and strict access restrictions. Consequently, my interaction with Git, and GitHub specifically, has been primarily as a user or consumer. Over the years, I have created many things that could be packaged as plugins in various languages. However, especially for JavaScript, which I know best, there was always a language barrier and a reluctance on my part to write documentation as typically proposed (I only write docs and comments when something isn't obvious from the code). So, now with the help of AI, I had a few days where I could steal some hours from work, a ready-made plugin, and I thought I'd give it a shot.

The code was ready and functional, but I wanted to rewrite it as a JavaScript class. I used four AI models for this: ChatGPT-4, DeepSeek, Qwen Code, and Kimi. ChatGPT was often the worst, but it was the best on rare occasions. The other three each had their own strengths, which I can now recognize, allowing me to direct the appropriate question to the right model (though, of course, that could change with the next version).

They were very helpful, especially in adding JSDoc comments and creating GitHub files like README.md. Their initial content wasn't great, but I corrected it while keeping their formatting. When the time came to create a TypeScript version, they were all terrible. I had never written TypeScript in my life, but I understood what it was and how it worked from reading some articles back when it was a hype. I knew more than the AIs did and had to point out every single issue.

They weren't much better at transforming my JavaScript "function" class object into an actual ES2022 class. The errors were sometimes hilarious, but they still helped me write the code 30% to 50% faster, as I only had to fix the parts they couldn't understand or review and point out the logical errors.

The real gain is that I previously knew nothing about all those specific GitHub files (like .gitignore, .gitattributes, etc.). I had probably seen them as a consumer but never paid any attention. Now I know what they are. I also now know how to write TypeScript, as most of the rewriting in that area was done by me because the bots couldn't grasp the newer versions of the language.

Well, without further ado, here is the link, and I await your comments:
https://github.com/ej-kon/DoubleRange

Edited by jkon
  • 2 Contributors
  • 3 Replies
  • 62 Views
  • 1 Month Discussion Span
  • Latest Post Latest Post by jkon

Recommended Answers

All 3 Replies

Member Avatar for jkon

it was ChatGPT 5 not 4 ... also it was KIMI 2

Member Avatar for WendyDolan

Wow congrats on your first plugin! That is awesome. I have been curious how much did AI actually help in creating it? Also does it work well with any fast web host or just certain setups?

Edited by WendyDolan
Member Avatar for jkon

Hello Wendy,

This is not my first plugin; I have been a programmer for more than 30 years now, and my company uses a client-side stack with plugins that I created or edited. This is my first on GitHub, as I haven't published anything publicly before. It was a learning experience needed for the next step that I am working on.

The main problem is with the documentation. I am not a native English speaker, so my documentation sounds a bit syntactically awkward. Although I speak English fluently (or so I think), when I write, you can clearly see the influence of my native language's syntax, which might sound strange to a native English reader. LLMs solved that problem for me. Also, in the README.md, bots helped a lot. No, they didn't write it—I did—but they proposed a skeleton and corrected my syntax.

They also helped a lot to make it look unified. My earlier versions had inconsistencies in formatting based on what was needed in each section. Their help with the actual code was minimal, but it was helpful. Especially, the confrontational tone of Kimi 2 made me rethink some code, and while Kimi 2 was wrong most of the time, in a few exceptions, it spotted some issues that I didn't (and neither ChatGPT nor Claude Sonnet did). Qwen Coder helped fill in the JSDoc that was minimal and made it really thorough. ChatGPT 5 was a big disappointment in that experiment. I will let it mature a bit and pay a month's fee to see if it behaves any better. It was worse than all the others in most cases, but on rare occasions, it provided some insights that others didn't. I will pay for its premium next month to test it further.

Also what do you mean by '''Also does it work well with any fast web host or just certain setups?''' ?

Edited by jkon
Be a part of the DaniWeb community

We're a friendly, industry-focused community of developers, IT pros, digital marketers, and technology enthusiasts meeting, networking, learning, and sharing knowledge.