I feel like sooner or later the standard for these types of discussions should become:
> "Here's a link to the commits in my GitHub repo, here's the exact prompts and models that were used that generated bad output. This exact example proves my point beyond a doubt."
I've used Claude Sonnet 4 and Google Gemini 2.5 Pro to pretty good results otherwise, with RooCode - telling it what to look for in a codebase, to come up with an implementation plan, chatting with it about the details until it fills out a proper plan (sometimes it catches edge cases that I haven't thought of), around 100-200k tokens in usually it can knock out a decent implementation for whatever I have in mind, throw in another 100-200k tokens and it has made the tests pass and also written new ones as needed.
Another 200k-400k for reading the codebase more in depth and doing refactoring (e.g. when writing Go it has a habit of doing a lot of stuff inline instead of looking at the utils package I have, less of an issue with Spring Boot Java apps for example cause there the service pattern is pretty common in the code it's been trained on I'd reckon) although adding something like AI.md or a gradually updated CODEBASE.md or indexing the whole codebase with an embedding model and storing it in Qdrant or something can help to save tokens there somewhat.
Sometimes a particular model does keep messing up, switching over to another and explaining what the first one was doing wrong can help get rid of that spiraling, other times I just have to write all the code myself anyways because I have something different in mind, sometimes stopping it in the middle of editing a file and providing additional instructions. On average, still faster than doing everything manually and sometimes overlooks obvious things, but other times finds edge cases or knows syntax I might not.
Obviously I use a far simpler workflow for one off data transformations or knocking out Bash scripts etc. Probably could save a bunch of tokens if not for RooCode system prompt, that thing was pretty long last I checked. Especially good as a second set of eyes without human pleasantries and quick turnaround (before actual human code review, when working in a team), not really nice for my wallet but oh well.
> "Here's a link to the commits in my GitHub repo, here's the exact prompts and models that were used that generated bad output. This exact example proves my point beyond a doubt."
I've used Claude Sonnet 4 and Google Gemini 2.5 Pro to pretty good results otherwise, with RooCode - telling it what to look for in a codebase, to come up with an implementation plan, chatting with it about the details until it fills out a proper plan (sometimes it catches edge cases that I haven't thought of), around 100-200k tokens in usually it can knock out a decent implementation for whatever I have in mind, throw in another 100-200k tokens and it has made the tests pass and also written new ones as needed.
Another 200k-400k for reading the codebase more in depth and doing refactoring (e.g. when writing Go it has a habit of doing a lot of stuff inline instead of looking at the utils package I have, less of an issue with Spring Boot Java apps for example cause there the service pattern is pretty common in the code it's been trained on I'd reckon) although adding something like AI.md or a gradually updated CODEBASE.md or indexing the whole codebase with an embedding model and storing it in Qdrant or something can help to save tokens there somewhat.
Sometimes a particular model does keep messing up, switching over to another and explaining what the first one was doing wrong can help get rid of that spiraling, other times I just have to write all the code myself anyways because I have something different in mind, sometimes stopping it in the middle of editing a file and providing additional instructions. On average, still faster than doing everything manually and sometimes overlooks obvious things, but other times finds edge cases or knows syntax I might not.
Obviously I use a far simpler workflow for one off data transformations or knocking out Bash scripts etc. Probably could save a bunch of tokens if not for RooCode system prompt, that thing was pretty long last I checked. Especially good as a second set of eyes without human pleasantries and quick turnaround (before actual human code review, when working in a team), not really nice for my wallet but oh well.