Code schizophrenia

Trying to create a new term.

So Cursor is doing an outstanding job of multi file edits, injecting instructions into our code request and using tools such as grep to grab code in our our project.

When things are working, its great. When they go off the rails, it’s difficult to recover. We have the limitations of the LLMs. Sometimes Cursor grabs way to many files, sometimes it doesn’t get the right one.

For example, this morning in composer I had the 2 files selected, one JSON, one python. Instructed it to simply put 4 hardcoded vars into the JSON and use that. Simple. < what i typed
It modified 5 files, over 100 lines of code changes, and took the main function of the app and replaced it with # put logic here !

This was an easy catch, I tried a few more time and just hand coded it.
Point is, when this happens it removes the reason to use coding assistant. Personally I can alt-tab, copy/paste code as fast as using a code editor, but project multi-file edits is where the goodness happens.

The glue (and value) is taming the LLMs! and context.

This will only become more relevant as the models get smarter.
Eventually the models will push back… try making code changes in a file with model = “gpt-4o” and you’ll understand what I mean.

I’ll leave the techy parts to you.
Good luck :slight_smile:

If you know what files to edit, I’d recommend using the normal mode in composer, not the agent mode. This will only use the context that you @.

Not only will this ensure the composer doesn’t edit files it shouldn’t be editing, but you’ll also tend to have a lower error rate due to the narrower context (more focused).