I have created 2 apps with 3 websites with Cursor but I am on the edge of giving up this IDE

Discussion
1

I want first to show my thanks to Cursor team. You have made a great tool for coding. It has totally changed everything in coding. but you can do better.

As we know, Cursor can use rule to guide the models. especially with agent model.

but when you are having conversations, the rules will soon or later be neglected. for example, you set the rule for minimum changes and no new files for fixing problem because it can cause infinite loop. but the llm model will forget it.

The worst part is llm model can try to delete your file or modify your files when you have rules telling the model not to.

If Cursor could not fix this problem, soon when another IDE supporting reasoning model with agent capacity and a good price at tokens, Cursor will lose its market for sure.

2

this is a fair problem, but I think you might face this on all IDEs.

context window limit is a limitation across all LLMs, and while Cursor can try their best to tackle it, it would be best to create new chats/composers whenever possible to keep the context sharp.

2 Likes
3

but the problem is , when to create a new chat/composer?

4

try to do modular smaller things with each chat/composer. if it’s starting to forget context, start a new chat/compsoer and mention @ composer history of prev composer

2 Likes
5

this tweet from @ericzakariasson conveys it better:

2 Likes
6

I don’t know if this will be helpful? My Best Practices for MDC rules and troubleshooting

Potentially prioritizing rules will keep it limiting the context that gets added?? Not sure.

7

So i’ve micromanaged my development and in a way treated the IDE AI as a dev team. You go through iterations. I actually build about 15 rules that I reference in chat to do a certain iteration and close the composer after. Think less big picture, use references and ask the agent to perform the action (smaller in scope). I think at the end of the day you can x10 but you need to manage iterations closely. I would recommend @ {rule} to help guide. I also use a filesystem reference, that I also update every time we add new folders-- to remind the system when they try to create new files.

8

one task = one chat

9

Is there a list of ‘best practices’ like this that is available?

10

thanks to everyone relying so far.

I agree that treating cursor as dev team. I am also doing it in similar approach.

I think we currently still need to use common software development approach. like to divide functions in individual models, each model share a entering parameters , so they can call each other.

however, it means we still need to spend more time on prod design than ever before. we certainly can use AI to acceclerate it.

11

I only wish Cursor was transparent about the current context limit in the chat.

When the context limit is reached, Cursor starts compressing/summarizing some of the information to stay within the context limit.

Currently we have no way to tell when this happens and have to guess based on the AI’s degraded response and forgetfulness.

3 Likes
12

not yet, but we’re working on creating guides for working with Cursor. happy to hear if there are any specific use cases you’re looking for!

1 Like
13

will Cursor support reasoning? Here is a project which is combining reasoning with Claude. It is much effective than just replying on Claude selft.
Would Cursor adopt the approach and introduce reasoning + Claude?