I think Claude trades quality for speed. From what I've seen it starts generating almost immediately even with a large token window. For smaller changes it is usually good enough, but larger changes are where I bump into issues as well. I'll stick to using change [somefunction] rather than change entire file.

I tend to iterate and limit the output by eg saying “make the smallest change possible, don’t rewrite the file, tell me what you want to do first” etc. It seems to respond well to change requests, with much apology. ChatGPT berates me about its own code and keeps saving shit I don’t want in memory, so I have to go back and clean up stuff like "Is testing Lambda functions directly in the AWS web console" and "Is working with testing and integration test fixtures in their software projects" when those are 2 of 100 things I'm doing. I'm using SAM for lambda, I might have run one in the console to bypass the API and now it's saved it as gospel. Half the benefit with LLM's is that they forget context when you start a new chat, so you can control what it focuses on.

What are you using to get your code into Claude. Continue, or something similar?

No I just drag files in, at least when I'm working with IntelliJ. Sluggish but works. With VSCode I've used Copilot but still break multi-file tasks out to Claude. I’m sure the eg Cursor workflow is better but it hasn’t bitten for me yet.

I also tried Replit and it felt amazing, but quickly got into a place it couldn't escape and it felt like it took too much effort to get it to change direction once it had committed to a plan. This was early, so it's almost definitely better by now.

I really like using aider (https://aider.chat) and heavily leverage the /ask command to have a quick chat on intent and context before prompting it for the code I need.

Aider feels like magic

is copilot improving behind the scenes ?

I assume you get the latest model, plus some goodies that GH is working on

Cursor is so good.