I noticed when chatting with the new agent with Claude 3.7 that it keeps adding and editing code for queries where I just ask it questions or ideas.
Anyone else notice this?
I even tried adding a rule to tell it to always first plan but it doesn’t follow it.
8 Likes
What always works for me is putting at the end: DO NOT CODE ANYTHING, JUST GIVE INFO. It has never ignored this.
3 Likes
right but i dont want to have to do it in every session 
2 Likes
Yes, I think its a problem with Sonnet 3.7.
I will disable it for a while.
Todays cursor is nuts.
1 Like
This is interesting, I was literally about to post here or on X that my last 20 responses from Cursor Claude 3.7 all started with verbatim “You’re absolutely right, …”
3 Likes
Tinker with the cursor rules (the general ones that are applied to all chats) to instruct the model not to code anything when you say [whatever keywords you want] or something in this spirit.
3 Likes
After disabling 3.7 things are MUCH better.
2 Likes
Hmm yeah I think 3.7 needs to get some guidance. By default it seems to try too hard to please you by fixing all sort of things and running the dev server and whatever. It’s a bit wild.
1 Like
use notepad “DO NOT CODE ANYTHING, JUST GIVE INFO OR UNTIL I TELL YOU TO CODE”
2 Likes
I noticed that you can change between ‘Modes’ mid-chat. So you can ask some question in ‘Ask’ mode and then switch back to Agent. This way forces a text-only response. It doesn’t access tools though so it won’t retrieve additional context (or so it seems). Maybe Edit mode is the answer here.
Either way, I’ve been finding 3.7 a bit harder to use so far. I might go back to 3.5 as well. I haven’t tinkered with my rules just yet though.
2 Likes
Yes, i have a shortcode for it. - aoni - analysis only, no implementation, which I have added to the rules.
So, aoni at end of messages. 
4 Likes
Ive had similar issues. 3.7 can end up going in circles more then 3.5
I have reverted to 3.5 for now.
1 Like
Yep, 3.7 (thinking, haven’t used vanilla too much) deadlocking itself in infinite loops. VERY annoying and burns out my premium requests :/. I’m switching back to 3.5 for now as 3.7 became kinda useless in the last day or so. Hoping the cursor team looks into / fixes it shortly!
2 Likes
Yes, the same thing happened to me.
When using 3.5, I like to use plan & act mode (Plan vs Act mode (xml prompt)), it can always think first, then get my confirmation, then start programming. But claude-3.7 always start changing code while thinking
1 Like
Btw, I had been using ‘Composer’ a lot previously. The new agent mode is more like previous ‘Composer’ than ‘Chat’.
Chat was safer - executed only when you told it to, so control was in your hands.
Compose - I found it more useful when designing early on in lifecycle, when new components had to be built quickly. So, it would help me build the base very quickly. But, when it came to already built/tested/working code, I always had to double check that it was not making sweeping code changes. 9/10 times, it did the wrong thing. So, I used ‘Chat’ mode for more controlled changes. My guess is ‘Agent’ is now 'Compose and ‘Ask’ is new ‘Chat’.
So, different modes help with different situations. These tools should give us guidance on what to use when, especially when we start new. or after a major upgrade e.g. use me for x,y,z and so on. @jake - a suggestion. Also, please confirm my understanding of the key difference between Agent, Ask and Edit modes. Thanks
1 Like
It’s nuts! I feel like it takes over when I code. Plz fix