Comment by perardi
2 days ago
That doesn’t reflect my (I would say extensive) experience at this point, nor does it reflect the benchmarks. (I realize benchmarks have issues.)
Are you using Claude as an agent in VSCode or via Claude Code, or are you asking questions in the web interface? I find Claude is the best model when it’s working with a strongly typed language with a verbose linter and compiler. It excels with Go and TypeScript in Cursor.
I have used it for GDScript, C++, Java, and other more general questions. Specifically, comparing it to other LLMs responses ESPECIALLY after incremental narrowing by prompt. Claud seems to randomly change approaches and even ignore context to the point you get the same circular issues you see in Copilot (do A because B is bad, then do B because A is bad or worse ignore everything before and do C because it's nominal). It seems more primitive in my sessions from the last time I used it (for a couple days) ~45 days ago.