It's working incredible for me. Opens up so many doors on so many projects. Added "picture in picture over fullscreen video" in one shot on one of my projects. One of the folks on my team does "@codebase where do you think <description of bug> is?" to find tricky bugs, and is having great success..!
I took a ~1200 line component and it broke it into ~5 separate components that all continued working without problems, pretty insane
Its working fine for me. On the first day it was acting really weird and doing a lot of things it should not, but now it seems to have improved since the last cursor update. The overall quality of the results are looking better, but at the same time it appears to always take a bigger approach than 3.5 for similar tasks.
@guilherme_kurtz I run into this too. I've even updated my cursor rules but without any luck. It ends up updating 15 files when I just want to solve a single step in the process.
@allnick One thing that seems to help me minimize those cases was reducing the context i gave to it. As i narrowed down the context and steps for him to work on, i got cases where he managed to get him to work just in a single file
@emad_ibrahim Can you imagine a step function difference from a foundation model update? Or do you think we're firmly in the regime of more incremental gains?
@rajiv_ayyangar it’s hard to predict anything any more. What’s mind blowing today is crap next week 😂. I love the pace but it’s hard to keep up.
I just love how efficient it’s been making me, I have been cranking out apps like no tomorrow. My latest one is an open source alternative to Motion - https://fluidcalendar.com/ once I figure out all the bugs, I will turn into a SAAS for a fraction of Motion (which I am a big fan of btw ❤️)
@rajiv_ayyangar what would be nice would be a constantly up to date model with the latest versions of packages and with a bigger context to understand the full project and not keep making the same mistake over and over or give me code for an older version of nextjs for example. Small annoyances compared to the 100x productivity I have gained. I know I know, I am spoiled :)
I’ve been using both for a bit now, and overall, I’d say Claude 3.7 is definitely an improvement. Some issues I had with 3.5 seem to be fixed, and it handles certain problems better. That said, it still makes a lot of mistakes, sometimes really simple ones, so don’t expect miracles.
One thing that did change is the introduction of the three modes: Chat, Edit, and Agent. All of them are pretty solid for their specific tasks. When using Agent mode with 3.7, you can actually see it analyzing everything more deeply and reasoning through tasks better.
@hussein_r You don't find it annoying to switch between the three modes? It seems like a nice intermediate step, but at some point I'd assume the model gets better at detecting intent and deploying the appropriate thinking mode.
@rajiv_ayyangar Yeah of course. Currently almost 95% percent of every request I make is with the Agent Mode. It does the best job and I doesn't need to switch between the modes. I hope that in the future, you would have just one mode, being great at everything, just as you described. That would be ideal.
More ticked off that cursor has removed Composer. Why? Chat not nearly as powerful and prone to annoying errors, wrong results. Bring back Composer, please!
It works great for me. I use 3.7 Extended Thinking to plan out what I'm going to do, and then regular 3.7 to execute the changes. It has been flawless.
I am really impressed with the outcome after switching to 3.7 from 3.5 Usage for sure is more but because it is getting accomplished more on its own! I hit 25 tool calls limit in a single prompt multiple times a day now so it makes sense also.
Replies
It's working incredible for me. Opens up so many doors on so many projects. Added "picture in picture over fullscreen video" in one shot on one of my projects. One of the folks on my team does "@codebase where do you think <description of bug> is?" to find tricky bugs, and is having great success..!
I took a ~1200 line component and it broke it into ~5 separate components that all continued working without problems, pretty insane
Its working fine for me. On the first day it was acting really weird and doing a lot of things it should not, but now it seems to have improved since the last cursor update. The overall quality of the results are looking better, but at the same time it appears to always take a bigger approach than 3.5 for similar tasks.
The Nifty
@guilherme_kurtz I run into this too. I've even updated my cursor rules but without any luck. It ends up updating 15 files when I just want to solve a single step in the process.
@allnick One thing that seems to help me minimize those cases was reducing the context i gave to it. As i narrowed down the context and steps for him to work on, i got cases where he managed to get him to work just in a single file
Graphify
@guilherme_kurtz Same here, first day was quite weird, but with the right prompting, you can see how powerful 3.7 is.
CuratedLetters
It’s working fine but so was 3.5. I can’t tell the difference but I love Claude with cursor. I just wish there was a more noticeable difference
Product Hunt
@emad_ibrahim Can you imagine a step function difference from a foundation model update? Or do you think we're firmly in the regime of more incremental gains?
CuratedLetters
@rajiv_ayyangar it’s hard to predict anything any more. What’s mind blowing today is crap next week 😂. I love the pace but it’s hard to keep up.
I just love how efficient it’s been making me, I have been cranking out apps like no tomorrow. My latest one is an open source alternative to Motion - https://fluidcalendar.com/ once I figure out all the bugs, I will turn into a SAAS for a fraction of Motion (which I am a big fan of btw ❤️)
CuratedLetters
@rajiv_ayyangar what would be nice would be a constantly up to date model with the latest versions of packages and with a bigger context to understand the full project and not keep making the same mistake over and over or give me code for an older version of nextjs for example. Small annoyances compared to the 100x productivity I have gained. I know I know, I am spoiled :)
Graphify
I’ve been using both for a bit now, and overall, I’d say Claude 3.7 is definitely an improvement. Some issues I had with 3.5 seem to be fixed, and it handles certain problems better. That said, it still makes a lot of mistakes, sometimes really simple ones, so don’t expect miracles.
One thing that did change is the introduction of the three modes: Chat, Edit, and Agent. All of them are pretty solid for their specific tasks. When using Agent mode with 3.7, you can actually see it analyzing everything more deeply and reasoning through tasks better.
Curious to see where they take this next!
Product Hunt
@hussein_r You don't find it annoying to switch between the three modes? It seems like a nice intermediate step, but at some point I'd assume the model gets better at detecting intent and deploying the appropriate thinking mode.
Graphify
@rajiv_ayyangar Yeah of course. Currently almost 95% percent of every request I make is with the Agent Mode. It does the best job and I doesn't need to switch between the modes. I hope that in the future, you would have just one mode, being great at everything, just as you described. That would be ideal.
More ticked off that cursor has removed Composer. Why? Chat not nearly as powerful and prone to annoying errors, wrong results. Bring back Composer, please!
CuratedLetters
It works great for me. I use 3.7 Extended Thinking to plan out what I'm going to do, and then regular 3.7 to execute the changes. It has been flawless.
I am really impressed with the outcome after switching to 3.7 from 3.5
Usage for sure is more but because it is getting accomplished more on its own! I hit 25 tool calls limit in a single prompt multiple times a day now so it makes sense also.
Absolutely love the performance and the output! Much better than 4o and the previous Claude versions.