Is C.4.S.T. broken or something?

Discussions
1

Spent many hours trying to solve a pretty damned difficult bug, and for hours I pressed on, just to end up somewhere even worse, with Claude Sonnet 4 Thinking. Then I rolled back and gave the task to Grok-3 and it just fixed it in matter of seconds.

Just you know.

2

It’s all good for me.

For maybe months now, I sometimes switch to Gemini, but never anything other than Claude
When it goes off track, I normally:

  • Check the source control view for the diffs to see if the code is as I want it.
  • Commit
  • Get it to write into a markdown file exactly what we have done in the chat. This is so I can resume in a fresh state.
  • I do sometimes do a reindex, but it is pretty fast at updating now.

Also, in my custom rules, it can access my docker container logs. My code hot reloads on changes, so it just checks the logs, looks at the errors.

I do also tell it to just do some analysis, drop a bunch of files into the context, and tell it to use it’s tools to do some “deep” sort of research into the issue. Not to fix it or doing any write operations. Also, if you have a UI, pass it alot of screenshots, including the url path. I’m not sure how you work with cursor though or the codebase, code, structure etc. Another thing is I’m starting to adapt the way my code is organised to work better with Cursor, so much smaller files. - but even recently large edits are becoming less buggy..

I think that’s something that needs to start shifting. I’m not sure if it makes sense, but codebases should now be designed to work with AI, not against it. AI driven development… So instead of like my code in folders like pages, components, hooks, etc, etc, I’ll have it in folders per feature, or “entity” of my app…

Up until recently when you dropped a folder into the context it only added metadata, like paths etc (I think?), so I removed all my folders and have a codebase with just files haha, so I could just select all and drop into the chat window. Now it’s just getting smarter so I’m really just chatting to it..

Sorry I’m not sure if any of this makes sense, it has been a long day… back to work…

2 Likes
3

Cool, thanks.

4

It sounds like a context limitation where it eventually forgets things. Best in such cases is to go to the prompt before it started going wrong and fix the prompt. Or start fresh chat with specific info.

1 Like
5

Speaking of, I find Cursor’s handling of too long context unfortunate.

Instead of removing just the earliest part of the conversation and replacing it with new prompts, they remove everything and generate a summary, like for the available new chat option, without a warning before or any information afterwards, other than the AI model thanking for providing a summary.

And we cannot see the actual content of this summary, we have to export the chat and read it outside of Cursor.

2 Likes