To vibecode or not, that is the question

Discussion
1

First, cursor is awesome. Great product. Thank you Cursor folks.

In the last weeks though, it looks like the Cursor prompting has been getting to eager to:

  1. Just generate lots of code, shoot first ask questions later, make it compile even if it does nothing like the prompt.
  2. Lost in the middle issues are becoming more and more frequent, even with simple prompts, such as, replace this constant with a state, it adds a ref.
  3. Claude 3.7 has become useless in Cursor, at least in our environment. It kind of stopped getting anything right a few weeks ago.

Some times, a bit of half baked code that has the right concepts is better than nonsense that compiles and does nothing of that it is supposed to do, and it looks like that is the current direction. I’m suspecting your backend injected instructions are getting too long, or complex, or contradictory (potentially with something the end user developer is trying to do).

What this is causing is having me and my team use cursor less and rely more on other tools or the API to generate code. Maybe a vibecoder mode would help the cause (that I can turn off), or having more knobs and switches to control what behavior we’re getting out of the agent.

Thoughts?

2

Sounds like you need to write much better Rules (in the AI settings).

3

My teams has a series rules we have been interesting for a while, since they made ‘.cursorrules’… and that is another thing, the have been having a diminishing impact and the agent now ignores them all the time. When we put the agent to write tests now, the agent forgets all the instructions after two loops.

Also, browsing around the forum and elsewhere it looks like there are a few people saying the same things.

Feel free to a look at the posts here that jail broke the injected context, it’s now massive, obviously the agent will forget/miss instructions.

1 Like
4

Look at that, they just patched " We’ve also fixed a long-standing issue where Always attached rules now persist across longer conversations. Agent can now also edit rules reliably." Thank you cursor team!

1 Like