There’s been a bit of debate whether Claude 3.7 or 3.7 Thinking is better for specific tasks.
For example, some people prefer non-thinking for frontend work, etc, but it might be a bit early to tell.
I’d like to hear everyone’s thoughts.
1 Like
maybe it’s a bit early, but when I use the think feature, it uses too many calls. i constantly get 25 call alerts, forcing the system to resume from where it left off. this causes the connections to break after a while. i think i get better results when i use the normal feature. if the call alert didn’t come too early, i would probably use the think feature.
With these latest developments, I think the team should increase the number of 25 call alerts. it should be at least 50 because 2 of the current new models are working very well and this causes them to work actively all the time. but this also causes call alerts to come.
2 Likes
agree with @mehmet-py the thinking model is too heavy for fast development
1 Like
I’ve been finding that the Thinking doesn’t integrate well into the agent. It is good for one-off requests and follow-ups, but I find that it very quickly loses track of my requests and starts making changes contrary to my request.
Thinking models seem to require a different kind of prompting, and giving them instructions that lay out an order of operations or steps they need to follow causes poor results. Perhaps the poor performance is because the Agent has carefully set prompts for a foundational model, and these prompts run contrary to effective ‘Thinking’ model prompts.
I’ve had good results with 3.7, its early days, but it feels the same as 3.5 to me.
DEFINITELY NOT AGREE WITH YOU DUDES! AS QUALITY OF CODE IT INCREASED A LOT WHEN I USE TO TRY ALL THINKING IN ALL STEPS! ITS GOOD TO THINK ALWAYS BEFORE MOVING NON THINKED MISTAKES! SO IT MUST BE OPTIONAL TO THINK IN EACH STEP! Keep your ideas to own!
One thing I noticed, 3.7 Thinking is worse at actions. It is just thinking and recommending some steps to debug etc. but that’s all, the final step is missing always. Anyone experiencing the same thing?
My experience with the thinking model is that it will tend to spiral out of control by conjuring up its own ideas and enhancements I didn’t ask for while it’s doing its work.
I ask for a simple “add X to Y”. It starts thinking and then starts working, and I see things like: “Now I’ll make X better by doing A, B, C and D…” then I’ll make A and B more efficient by adding E, F, G and H…
And so on, as it quickly spirals out of control, adding features and optimizations never requested, completely forgetting the narrow scope of the original request and then hitting the tool limit and leaving its work unfinished.
I have a rule in my .cursorrules file that I’ve used for ages that is basically:
MAKE ONLY THE MINIMUM CHANGES REQUIRED TO ACCOMPLISH YOUR TASK. DO NOT MAKE ANY IMPROVEMENTS, REFACTORING OR CHANGES THAT ARE NOT EXPLICITLY REQUESTED.
It helps a bit, but not with 3.7. Maybe its not reading my rule 
Yeah, I usually don’t have this issue with 3.5, it has always done a pretty good job of addressing everything I ask for, but 3.7 is a little bit more over-zealous if you let it, so I’ve been including a similar instruction with each prompt 
3.7 and my views on the thinking mode:
thinking mode doesn’t work as I expected. it has problems with context building. it forgets messages a few messages ago. it has difficulty understanding what is being said. it misunderstands. it is very difficult to build projects!
in 3.7 normal mode, I think there is a problem with the model added to cursor. because as far as I know, when the cursor team integrates claude into cursor, they integrate it in the logic of the standard model of the cursor platform. they add different features. but there are some problems while doing this.
without designing the project, without context, without understanding what is being said, this model also has difficulties. both have almost the same problem.