fix: Ensure consistent language in coder prompt descriptions

[fry69]

fix #1850

Thanks to @businistry for reporting and @jorgecolonconsulting for digging into this and providing the fix!

[jorgecolonconsulting]

you're welcome!

[paul-gauthier]

Why do we think this rephrasing "works better" in general? It seems quite similar.

[fry69]

Why do we think this rephrasing "works better" in general? It seems quite similar.

Mystery of LLMs I guess? There is a slightly different meaning or instruction logic the new wording, I think.

[jorgecolonconsulting]

Why do we think this rephrasing "works better" in general? It seems quite similar.

Not so much “think” as in actually tested against a couple of different models. The biggest thing it fixes is where the default strong model (gpt-4o) that Aider uses responds in Spanish as many have reported here. That leaves a bit of a bad first impression when you’re expecting Aider to respond in your expected language.

LLMs are weird in that slight changes in a prompt can have dramatic changes in behavior for unknown reasons. Even if they are essentially saying the same thing. I’m sure you’ve experienced that yourself.

Have a look here where I was able to reproduce the problem and where this prompt did better: #1850 (comment)

[jorgecolonconsulting]

A thorough test across many models can make sure that prompt did better. We could even ask the issue reporters that experienced responses in Spanish what their chat history and incorporate those into the tests.