I came here to post about the same issue:
I think prompt engineering for this is the wrong approach. Prompt engineering is hard enough as is and code linters/formatters have been working well for many years already.
What seems to be the issue right now is that the linting configuration Cursor uses for the “Iterate on lints” feature is not configurable. Ideally it would intelligently check the current repository for files like
- .commitlintrc
- .pyproject.toml
- .gitlint
- etc. (there are many possibilities here depending on the programming language/framework you’re working with)
and then surface those linting/formatting violations to Composer (normal and agent). This would do wonders to improve model output and drastically reduce time spent fixing errors downstream.
In addition, Cursor should probably check for any linting/formatting configuration on user/workspace/default settings level of the editor itself like some language server extensions allow.
I realize this feature is still in beta due to not being trivial to implement… but the current implementation (as of Cursor 0.44.8) is already so great I wish it were fully implemented already because it’s so ■■■■ useful.