Raw LLM Responses

It takes some time to get used to AI coding agents. Claude seems better at going for higher quality code I think, being a bit more proactive rather than doing a very minimal MVP of whatever you asked for... but even if you use Claude Code it really helps to be specific about these things if you want it to avoid 1000 line classes, to make it all modular, to avoid reusing code and so on. My experience with ChatGPT and Codex CLI is... usually not so good. It should be better now, but I need to do better tests I think. But with Claude Code and Copilot CLI, using Claude LLMs, I have gotten some good results. Made it write some MD files about the structure of some command line tools I had made for stuff like loading CSVs into a database, export tabels to CSV, do FTP stuff and so on... and in this md file there were instructions on the structure of these projects and I have then made new projects in the style, but with different datasources and such. For some simple apps they might take under an hour where most of that is just the LLM doing stuff and me doing other stuff, and with more complicated things I need to do more for it to get it right. But AI agents can save people a LOT of time as it is now, and they are only getting better. AI cannot replace developers though. Problem is the things we do, we do for people who do not know what they want... the spec that the AI needs is not what non-coders would give the AI. But each developer certainly can become a lot more productive with AI. There are still a lot of work that is not just writing the code itself and that is to f.ex. figure out what a company or person in a company actually needs, describe workflows, figure out how things actually happen.