Claude is on another level now.
Can do amazing things :D.
I’ve been having ChatGPT write functions and troubleshoot for me for a while- working in a chat window and I copy/paste things into VSCode.
I’m curious about trying out one of the models in a more agentic fashion where it’s creating files and writing code on its own. What’s the best way to get started with that?
I’m in a similar camp. Used it (and Claude) for real projects, backend + frontend + infra, and it’s a massive productivity boost when you know how to frame the problem.
It’s great at scaffolding services, wiring APIs, designing schemas, and even tuning queries from real logs. That said, it’s still not a ‘drop in and forget’ tool for complex, multi-module systems. You have to guide it and review everything. Treat it like a very fast junior/senior hybrid who never gets tired.
2 Likes
@alexsam986 thanks for this, what do you think of each one’s Lua coding ability?
My results here have been a mixed bag.
Sometimes I’m pleasantly surprised by the approach, and coding, other times the things (both ChatGPT, and Claude) fall hard on their face, and if I’m not careful checking both’s code I can waste hours going down a rabbit hole, and as you mention this usually happens for complex multi-module systems.
I’m giving as much input for scaffolding as possible for each platform (e.g. Files, Instructions, Memory), and I’m as careful as I can be with inputs / prompts. H. Rickover “so clear that it can’t be subject to interpretation”.
Both are pretty good at both SQL (I’m using MariaDB), and JavaScript (and JQuery).
It really is extending my grasp.
One of my past laments when writing an app is you always have to keep your output something that one guy can do. Today I give it credit for being almost like having a junior programmer (albeit a junior programmer with a drug problem). I can usually tell when the apps start hallucinating before I get into too much trouble.
I don’t think that the AI thing is going away so I’m committing to coding with it as a partner. I’ve even got some some Linux like commands stored so that I can do things like enter ARW (Analyze, Report, Wait for instructions), and it will do just that.
Please share any tips if you have time!
Thanks again.
I find that the more complicated the task, the better an overall job it does. But that could be because doing 95% of a big complicated task feels like a good outcome, and for me to only do the remaining 5% isn’t a big deal.
On the other hand it often makes an absolute mess of simple problems, to the point that I have to undo everything and just do it myself. Yesterday I noticed that it kept asking me where functions where, and would write some code in the chat section then ask me to paste it into the code myself. Claude had somehow forgotten that it already had access to the files since it was running in the VSCode extension…