I just have to give you major props for your videos! They’re super on point, no unnecessary rambling, and they show exactly what you need to know. Plus, your way of explaining things is awesome! Clear and concise, yet still diplomatic. Absolutely great content. Keep it up! I tried both Cursor and Windsurf for different projects. Can't decide yet which one is better. Both have their ups and downs. Also read one of your comments and Aider is indeed a tool that I wanted to try out, too. Would be really great to see in a video what you think about it (esp. compared to Cursor and/or Windsurf)!
POWER TIP. Ask agent to do analysis only. No code. Somehow you can fix many issues if you ask it to solve problems in small chunks. My flow: Fix it -> Fix it, here is a doc -> what could be the issues? No code. -> what code is doing X? -> here is the code snippet, why we can have issue X? -> explain in pseudocode -> how can we debug X? No code. -> o1 preview help -> let's get a few other ideas how to inplement X -> zoom out. No code. So far the main work I have to do is to manage context, know when to provide doc, when to use vector codebase, what files to link, how to prepare info by providing comments and pseudocode, how to debug, or ask provoking questions, limit task size to be digesteble, select model that will solve the issue.. a bit like a quest..
The main problem: model performance degrades a lot after 3000 tokens/300 lines of input code. So sometimes I go directly to claude client outside cursor, as at least it works without many rules overloading context window.
Totally agree. If you’re using compose, commit your code often. Everything was going great and then all of a sudden it’s just carpet bombed my project that has nothing to do with the piece I was working on. Use with care! 5:085:09
Yep 100% - and because it takes so long to run, you're kind of left spinning your wheels going back to the checkpoint, retrying, and hoping for the best. This is why I still use Chat most of the time.
@@VoloBuilds Yes, I agree. I actually used composer a lot with the frustration coming along, but you have shifted my mind to rely more on Chat, which I actually neglected a lot. Thanks for your valuable input and channel!
Yes, the main problem with using the Composer agent is its tendency to overwrite existing code, replacing it with it's generated lines. Reliability is crucial, and I often find myself spending more time rewriting and reviewing its output than if I had coded it manually without AI. That said, Windsurf remains the best AI code editor available right now.
i would love to see someone from the cursor team respond to this’s video as it makes some great points and would be really helpful to get their perspective and any tips for addressing some of the apps limitations in the short term before fixes are available.
@ You’re welcome! Either way is ok. Most people call me Jay but for some reason I always sign my messages with Jason so it confuses people. Let’s stick with Jay and I’ll be sure to do the same lol
You don't need to commit all the time when using the composer agent. It makes checkpoints at each chat, so you can just revert to how the code looked before if you don't like the most recent output. That should help with the content-removal issue. You can find the button for reverting to the checkpoint at the top of your most recent message to the composer agent.
Yeah this is a great tip and definitely helps, but I found that it will often still descend down into code deletion - and the annoying thing is having to wait for it to regenerate the code, especially if using agents and relying on it to find/read files (takes even longer)
Great video, thanks! I noticed when I switched on the beta feature - lint errors my composer basically crushed and stopped functioning normally. At beggining removed the linter errors but there was no effect in coding. I switched off the feature and composer works again nicely. When I create new composer I pass over a prepared description/notebook, start as regular composer and tell to check @codebase and than switch to agent. Also additional .cursorrules specifying not to delete code before verifyin, etc is improving the performance.
Absolutely! I have a page describing the evaluation criteria in detail here: volobuilds.github.io/VoloScore/ and I will be adding the leaderboard section soon. Good opportunity to use one of these tools to build it as you suggested :)
Can you tell it to iterate on it's code until a test succeeds? Would that be a good use case for agents ability to use tools? Assuming that it even worked.
Yeah, this latest update sounds like it was rushed out in a mess which is a shame. I haven't upgraded as a result because it just sounds like things got worse rather than better. Shows why it is good to have competition be it from windsurf or copilot. If someone can figure out how to get something near as good as cursor tab autocomplete, they will then finally have some proper competition.
Yeah - I'm actually sticking with the more reliable chat flow for now - which thankfully is still an option. And yeah, the auto-tab in Cursor is next-level for sure!
i tried so many times to use composer but i always endup using chat. Mostly because chat also makes me learn my own code. Love your content . keep pushing ✌🏽
Very interesting, since I have much different experience with composer agent - almost flawless experience, thoughtful code etc. MUCH different from old composer, which was quite unpredictable (magic or mayhem).
@@WatsonCyrusAnikwai I think Cursor Chat is the most reliable/fastest option. Windsurf is pretty good and better than compose at finding relevant files (in my experience) but slow and I heard also suffers from code deletion (though not as much as compose)
Personally my main is Cursor Chat right now because of the iteration speed and reliability. I don't mind tagging the files all that much if it means it starts generating code immediately. Also the auto-tab is very useful for me as I get into the weeds of the code I'm writing.
Yea makes sense - although I will say it is still possible to keep using the Chat as it was before, which is more stable. That's what I'm doing for now.
It's been working very well for me, definitely some errors to work out but the amount I can get done in a large project is pretty fantastic. I'll probably stick with the current version that came out this past week until I see a review for a new version that has major improvements
That's great - let us know if you have any tips for getting the most out of it! Particularly when it comes to stability with the file changes and removing large portions of code.
Great vid. It is super useful for new features that require making a lot of files. Run it, then do not proceed but let it fix the errors. Only then proceed. If it can’t, discard all changes on Github and try with Windsurf. Proceed with Windsurf the same way, eventually discarding and changing. Once all is set up, switch to more reliable Chat. As of this update, I find Composer useful. Also lint errors fixing is a great step in the right direction. Aide, meh. Anyone likes using that?
Hahaha yes that workflow sounds very familiar 😂 I'm mostly staying on Chat for now because of the stability. Have you tried Aide? I was thinking of reviewing it - curious to hear what you think. Any other tools you've enjoyed using?
@VoloBuilds I've tried alot of local llms. nothing comes even close to claude. deepseek 236b is not bad. but you can't run that locally. so you still need an api. and i prefer the one time payment, not paying as i go. much more manageable. windsurf still the one for me in agentic stuff.
Yeah agreed - nothing comes close to Sonnet 3.5 - and personally I don't even see that much value in local llms over cloud ones. Maybe eventually if they get good it will improve speed? But then running them on GPUs in the cloud would probably be even faster.. idk haha
I tried windsurf and also cursor. there are a bunch of flaws in cursor, where is the documentation ? how do i know how to use the IDE and the different features ? why in composer it writes the code and also changes the actual code in the files on the left it's distracting. in Windsurf you see the code changes only on the left side. the pricing of cursor is higher, why ? it's the same service. TBH windsurf wins in my opinion.
I did a pretty thorough comparison of the two in a separate video and each has their strengths/weaknesses but this update doesn't really move the needle for Cursor IMO
I’m blown away at the composer hate. I loathe the chat but love the composer. One minor tip for less “deletion” is to update cursor rules to be explicit ;)
@@cyberpsilosis Cursor has a special feature where you can create a '.cursorrules' file and each interaction with the AI will abide by these rules and take them into account. Useful for things like telling the AI to write in a certain style or use certain frameworks and conventions. Sounds like Bryan found a way to use that to improve composer quality
@@VoloBuilds I find the chat has way too much output slop. Long conversations instead of just getting to work on what I ask. I don’t want to have a convo usually I almost always want the model to take action. Chat also weirdly limits itself in file creation and modification. I feel like it’s always telling me one file at a time where my personal experience with composer is much more holistic changes to my code base. One of my cursor rules is “when you’re talking about making changes to code files, never ever say things like ‘the rest of the file unchanged’. Please always be explicit about all the changes and retained code”
Not sure if you saw it, but I made a video trying out Windsurf: ruclips.net/video/ouT8ItvJla8/видео.html - if you wanted to see more, let me know what you're curious about!
@ I moved to bolt new from cursor but now am amazed by windsurf. You can literally build production grade full stack applications with it. You can search on YT to see what people are building . I am made a a supabase template if you want to quickly build a full stack website with it
Compose has been shit since the start.... I never understood the hype around compose. Just having a single interface of chat is way easier to understand and use...
Yeah I'm not sure why there would be two interfaces for this as it feels like just one function. In my experience, compose has that code deletion/laziness issue and chat does not. No idea why they would be different. Hoping Cursor can tune it up and standardize the interface to avoid confusion for new users
Hello Volo, I am just reaching out to see if you can help us with an app development. How can we setup a time to meet and talk about your rates? Please let me know. I sent you an email.
Hey there! I am currently at full capacity for work, but will re-evaluate in a few months. Reach out again if you're still looking for help at that time. I am likely going to also create a community for builders in the next few months so perhaps someone else can help you out too!
I just have to give you major props for your videos! They’re super on point, no unnecessary rambling, and they show exactly what you need to know. Plus, your way of explaining things is awesome! Clear and concise, yet still diplomatic. Absolutely great content. Keep it up!
I tried both Cursor and Windsurf for different projects. Can't decide yet which one is better. Both have their ups and downs. Also read one of your comments and Aider is indeed a tool that I wanted to try out, too. Would be really great to see in a video what you think about it (esp. compared to Cursor and/or Windsurf)!
This absolutely made my day :) thank you very much!
I am excited to do more reviews! Deciding between a few options right now
POWER TIP. Ask agent to do analysis only. No code. Somehow you can fix many issues if you ask it to solve problems in small chunks.
My flow: Fix it -> Fix it, here is a doc -> what could be the issues? No code. -> what code is doing X? -> here is the code snippet, why we can have issue X? -> explain in pseudocode -> how can we debug X? No code. -> o1 preview help -> let's get a few other ideas how to inplement X -> zoom out. No code. So far the main work I have to do is to manage context, know when to provide doc, when to use vector codebase, what files to link, how to prepare info by providing comments and pseudocode, how to debug, or ask provoking questions, limit task size to be digesteble, select model that will solve the issue.. a bit like a quest..
The main problem: model performance degrades a lot after 3000 tokens/300 lines of input code. So sometimes I go directly to claude client outside cursor, as at least it works without many rules overloading context window.
Great tips, thanks for sharing Yurij! Good reminder to take it in small chunks.
@@yurijmikhassiak7342 let’s go for a beer 😂
Totally agree. If you’re using compose, commit your code often. Everything was going great and then all of a sudden it’s just carpet bombed my project that has nothing to do with the piece I was working on. Use with care! 5:08 5:09
Yep 100% - and because it takes so long to run, you're kind of left spinning your wheels going back to the checkpoint, retrying, and hoping for the best. This is why I still use Chat most of the time.
@@VoloBuilds Yes, I agree. I actually used composer a lot with the frustration coming along, but you have shifted my mind to rely more on Chat, which I actually neglected a lot. Thanks for your valuable input and channel!
@@innovatune_ai That's awesome - thanks for sharing and I'm glad that this helped!
Carpet bombed, haha!
Yeah, things develop quickly. Step size, git often, test often.
Yes, the main problem with using the Composer agent is its tendency to overwrite existing code, replacing it with it's generated lines. Reliability is crucial, and I often find myself spending more time rewriting and reviewing its output than if I had coded it manually without AI. That said, Windsurf remains the best AI code editor available right now.
It does happen to Cascade too, need to preemptively prompt about not altering existing logic. Had to rollback to known good commits a few times.
ctrl z that shit bro
Luckily I save versions with git and emphasized several times in .cursorrules not to remove any existing code 😂
i would love to see someone from the cursor team respond to this’s video as it makes some great points and would be really helpful to get their perspective and any tips for addressing some of the apps limitations in the short term before fixes are available.
Great video, Volo! Thanks for all you info, I really appreciate it! Jason
Thanks Jay :) btw do you prefer Jay or Jason?
@ You’re welcome! Either way is ok. Most people call me Jay but for some reason I always sign my messages with Jason so it confuses people. Let’s stick with Jay and I’ll be sure to do the same lol
Haha alright sounds good :) glad you are enjoying the videos and thanks for your support!
You don't need to commit all the time when using the composer agent. It makes checkpoints at each chat, so you can just revert to how the code looked before if you don't like the most recent output. That should help with the content-removal issue.
You can find the button for reverting to the checkpoint at the top of your most recent message to the composer agent.
Yeah this is a great tip and definitely helps, but I found that it will often still descend down into code deletion - and the annoying thing is having to wait for it to regenerate the code, especially if using agents and relying on it to find/read files (takes even longer)
Great video, thanks! I noticed when I switched on the beta feature - lint errors my composer basically crushed and stopped functioning normally. At beggining removed the linter errors but there was no effect in coding. I switched off the feature and composer works again nicely. When I create new composer I pass over a prepared description/notebook, start as regular composer and tell to check @codebase and than switch to agent. Also additional .cursorrules specifying not to delete code before verifyin, etc is improving the performance.
Please build a simple app where you display the detailed VOLO scores of all tools you've tested
Absolutely! I have a page describing the evaluation criteria in detail here: volobuilds.github.io/VoloScore/ and I will be adding the leaderboard section soon. Good opportunity to use one of these tools to build it as you suggested :)
Can you tell it to iterate on it's code until a test succeeds? Would that be a good use case for agents ability to use tools? Assuming that it even worked.
Yeah, this latest update sounds like it was rushed out in a mess which is a shame. I haven't upgraded as a result because it just sounds like things got worse rather than better. Shows why it is good to have competition be it from windsurf or copilot. If someone can figure out how to get something near as good as cursor tab autocomplete, they will then finally have some proper competition.
Yeah - I'm actually sticking with the more reliable chat flow for now - which thankfully is still an option. And yeah, the auto-tab in Cursor is next-level for sure!
i tried so many times to use composer but i always endup using chat. Mostly because chat also makes me learn my own code.
Love your content . keep pushing ✌🏽
Thanks Baniya! Appreciate the support. Totally agree re: using chat!
Can you share some examples of prompts that made Cursor Composer delete files and lines like you are saying?
It seems pretty random but happens more often when dealing with a complex feature and/or a lot of context.
Very interesting, since I have much different experience with composer agent - almost flawless experience, thoughtful code etc. MUCH different from old composer, which was quite unpredictable (magic or mayhem).
I run into the same issue as well where it keeps on deleting bunch of code. I stick with chat for now until this is improved.
Yup, same! Hope they can figure out what is different between the two and fix it
@VoloBuilds is windsurf doing it better? I might consider trying it.
@@WatsonCyrusAnikwai I think Cursor Chat is the most reliable/fastest option. Windsurf is pretty good and better than compose at finding relevant files (in my experience) but slow and I heard also suffers from code deletion (though not as much as compose)
@@VoloBuilds thanks man - appreciate your reply.
great timing I have just got access to agents:) - thanks
Hey micbab! Nice, let me know how it goes for you
are you sticking to windsurf for now?
Personally my main is Cursor Chat right now because of the iteration speed and reliability. I don't mind tagging the files all that much if it means it starts generating code immediately. Also the auto-tab is very useful for me as I get into the weeds of the code I'm writing.
The overriding issue is the model not cursor use Claude.
The weird thing is that the problem doesn't happen when I use Cursor Chat rather than Cursor Compose 🤷
Great analysis! For now, I'm just sticking to the old release until all these problems get fixed :)
Yea makes sense - although I will say it is still possible to keep using the Chat as it was before, which is more stable. That's what I'm doing for now.
It's been working very well for me, definitely some errors to work out but the amount I can get done in a large project is pretty fantastic. I'll probably stick with the current version that came out this past week until I see a review for a new version that has major improvements
That's great - let us know if you have any tips for getting the most out of it! Particularly when it comes to stability with the file changes and removing large portions of code.
Composer deleting entire blocks of code and then gaslighting you that it didn't is one of the crazy moments of AI.
Yeah this is a big problem!
By the way, the lint iteration feature was already available to me in version 0.42
Oh wow, I assumed it was a part of the agents. Interesting!
Great vid. It is super useful for new features that require making a lot of files.
Run it, then do not proceed but let it fix the errors.
Only then proceed.
If it can’t, discard all changes on Github and try with Windsurf.
Proceed with Windsurf the same way, eventually discarding and changing.
Once all is set up, switch to more reliable Chat.
As of this update, I find Composer useful. Also lint errors fixing is a great step in the right direction.
Aide, meh. Anyone likes using that?
Hahaha yes that workflow sounds very familiar 😂 I'm mostly staying on Chat for now because of the stability.
Have you tried Aide? I was thinking of reviewing it - curious to hear what you think. Any other tools you've enjoyed using?
@VoloBuilds I've tried alot of local llms. nothing comes even close to claude. deepseek 236b is not bad. but you can't run that locally. so you still need an api. and i prefer the one time payment, not paying as i go. much more manageable. windsurf still the one for me in agentic stuff.
Yeah agreed - nothing comes close to Sonnet 3.5 - and personally I don't even see that much value in local llms over cloud ones. Maybe eventually if they get good it will improve speed? But then running them on GPUs in the cloud would probably be even faster.. idk haha
@ same
It's funny windsurf is kinda like Claude. The littler guy but actually the better guy.
very good analysis!!
Thank you, Himanshu!
I tried windsurf and also cursor. there are a bunch of flaws in cursor, where is the documentation ? how do i know how to use the IDE and the different features ? why in composer it writes the code and also changes the actual code in the files on the left it's distracting. in Windsurf you see the code changes only on the left side. the pricing of cursor is higher, why ? it's the same service. TBH windsurf wins in my opinion.
I did a pretty thorough comparison of the two in a separate video and each has their strengths/weaknesses but this update doesn't really move the needle for Cursor IMO
I’m blown away at the composer hate. I loathe the chat but love the composer.
One minor tip for less “deletion” is to update cursor rules to be explicit ;)
Wow interesting! Can you share what you prefer about it and what sort of cursor rules you have that help the code deletion?
What rules are you referring to?
@@cyberpsilosis Cursor has a special feature where you can create a '.cursorrules' file and each interaction with the AI will abide by these rules and take them into account. Useful for things like telling the AI to write in a certain style or use certain frameworks and conventions. Sounds like Bryan found a way to use that to improve composer quality
@@VoloBuilds tysm! I would love to learn more about this. Seems that would be the place to fix the composer issue
@@VoloBuilds I find the chat has way too much output slop. Long conversations instead of just getting to work on what I ask. I don’t want to have a convo usually I almost always want the model to take action. Chat also weirdly limits itself in file creation and modification. I feel like it’s always telling me one file at a time where my personal experience with composer is much more holistic changes to my code base.
One of my cursor rules is “when you’re talking about making changes to code files, never ever say things like ‘the rest of the file unchanged’. Please always be explicit about all the changes and retained code”
They shipped new update without proper testing
YOLO 😂
I wonder if they felt the pressure from Windsurf or if it was just a coincidence
I finally got the automatic update
Nice - let us know what you think about the agents!
Hiii you should make a video about windsurf agent new workflow. It’s amazing 😢
Not sure if you saw it, but I made a video trying out Windsurf: ruclips.net/video/ouT8ItvJla8/видео.html - if you wanted to see more, let me know what you're curious about!
@ I moved to bolt new from cursor but now am amazed by windsurf. You can literally build production grade full stack applications with it. You can search on YT to see what people are building . I am made a a supabase template if you want to quickly build a full stack website with it
It sounds like Cursor Agents are just like Junior devs :D
If you replaced every "Agent" with "Junior Dev" in your script, it would still make sense 💀💀💀
lol! This is actually true 😂😅
Great review
Thank you! More coming soon
Compose has been shit since the start.... I never understood the hype around compose. Just having a single interface of chat is way easier to understand and use...
Yeah I'm not sure why there would be two interfaces for this as it feels like just one function. In my experience, compose has that code deletion/laziness issue and chat does not. No idea why they would be different. Hoping Cursor can tune it up and standardize the interface to avoid confusion for new users
I had a non-productive day today because of the new updates 😢
Dang - well at least we can still use Cursor chat! And don't worry, it's good that you are trying new things - that habit will take you far!
Hello Volo,
I am just reaching out to see if you can help us with an app development. How can we setup a time to meet and talk about your rates?
Please let me know. I sent you an email.
Hey there! I am currently at full capacity for work, but will re-evaluate in a few months. Reach out again if you're still looking for help at that time. I am likely going to also create a community for builders in the next few months so perhaps someone else can help you out too!