I’ve been a user of CursorAI for a week, it’s incredible. I’m currently on the CursorAI pro version $20/month. It’s currently using Claude 3.0, however I was under the impression it was updated to Claude 3.5. Claude is in March 2024.
I’m looking to enhance my setup, I’m a mid-lvl engineer at best, probably akin to junior.
What are your experiences and recommendations? I’m reading that O3 mini is powerful however some people are experiencing a drop off in productivity.
Any guidance on advancing my setup would be appreciated and I’m sure new comers would find this useful.
honestly, claude 3.5 sonnet is still the best for coding, but it does lose context a lot of times, so the workflow that works for me is:
create a requirements.md and completed-tasks.md file using the latest openai model (they’re pretty good with content generation for any case) and then give those files to sonnet in composer and go ahead with that
when sonnet gets stuck somewhere in a loop, i try to start a new chat or use the deepseek r1 model to tell me how to fix that issue and then continue with sonnet
i did try o3-mini and its good with code too, but still not as good as sonnet
This was helpful, thank you. I typically speak to chat exclusively and write in my requirements on the go. Having a requirements approach sounds like the right lvl up for me.
Also for claude 3.5, I still don’t understand why my claude keeps telling me it’s 3.0. I’m on the $20 pro version, is there something i need to do to update?
claude 3.5 sonnet Currently, the coding ability is still stronger. You can use ChatGpt to generate documents, project designs, and basic rules (it is recommended that cursor officials add the function of automatically generating rules later). However, claude still intermittently has an infinite loop. At this time, I choose to use ChatGpt to process the segmented code (don’t give him the whole code, otherwise it is easy to have an infinite loop)
Deepseek sometimes misunderstands the meaning, whether in Chinese or English. I feel that the knowledge base is not that comprehensive at present, but there are better prompt words
First select [save] to test whether there is an error (if automatic saving is enabled, test directly). If there is no problem, use [Accept]. If there is a problem, adjust it. If the code is modified randomly or the previous function is deleted, select [Reject]. Note that when using the cursor, you must use git to save continuously. For the same function, your cursor may create multiple files or duplicate code, and it may not delete the file when it determines that an error needs to be deleted. It may also delete the completed file inexplicably. The context problem is still a problem that AI cannot avoid, and it can only be solved by constantly adjusting the cursor rules. GPT is better for local code and small functions.
have you tried going to cursor settings and checking which models you have enabled? they dont enable newer models by default so you’d have to do that manually if you’ve been using cursor for a long time
Thanks for that suggestion, I went into settings and 3.5 is checked off but when I ask the AI which model it’s using, it tells me it’s Claude 3 Sonnet, perhaps that’s just a response indicating it’s a derivation of Model 3.
I would look into which of these models is the latest claude too. One of them is going to be much better. Anthropic has done a bad job of naming their recent models (typical software industry stuff) and so the newest 3.5 sonnet model is called 3.5 sonnet (latest). Yup. Not great. But in the screenshot you provided, I have my claude 3.5 disabled/unchecked and I have the claude 3.5 sonnet (20241022) checked. I think that is the newer smarter one that is much better at coding. I could be wrong but that’s what I think from when I looked into their messy naming. I don’t seem to see that claude model listed in your screenshot though. Others may be able to verify for you based on how their one looks. I do have some custom models I added a while ago so I can’t say for certain if it should appear in your list. Also there is a setting to extend context in cursor settings. I have that enabled. While I’m here, you will need to invest more into improving your cursor workflow/setup as you use it more and more. To make it perform better. Setting up the correct rules. Having it write documentation and then using that documentation to generate better code. Following certain coding conventions… You can’t rely on the “.cursorrules” file either as it often ignores it. I normally “@” the file in the prompt. But I also have to state to follow the “@rules.md” for instance in the prompt, or else it will ignore the “@rules.md” in the prompt. Also, I agree with others that claude 3.5 sonnet (latest) is the best at coding. When stuck, the thinking models (r1, o1, o3-mini…) can come in handy. But the thinking models are not practical and often not good at writing code. r1 just loves to think and has a tiny context window, so it won’t be able to edit your code or write code for your app accurately in multiple instances. I’m starting to use MCP servers now and I’m experimenting with that, hoping to get better results. I would also caution against using deepseek v3 unfortunately as all it has ever done for me is create bugs and destroy the codebase in any way it can. Your time, energy, focus and mood are worth more than what v3 will save you financially. I wish this wasn’t the case because it’s so fast and cheap. For me, v3 creates more problems than it fixes. It’s also hyperactive and doesn’t follow instructions that well (for me). Any model other than claude or maybe o3-mini is a waste of time. Ignoring the hype is what makes us intelligent. Its what makes us claude
Same here… Claude is significantly better than o3 mini high… as for gemini 2 pro, i tried it and beyond the fact that it has a temper, i believe it to be even worse than o3 mini high. However large memory
with business account and can have access to 03 mini high, stovk using claude 3.5 sonnet 20241022, also have never been able to get yolo to work. just wondering if anyone can recomend something
