Cursor vs Cline | 240k Tokens Codebase Side-by-Side AI Coding Battle

[@MarvijoSoftware]

Hi. I'm glad to announce that you can now become a member of the channel to receive exclusive members-only videos, early access to upcoming videos, personalized content and more. Join here: kzbin.info/door/e3Zd8h8h9HYHWJva8DvrlAjoin

[@puremajik]

Thanks I really appreciate the rigorous evaluation. I would love to see windsurf vs cursor

[@MarvijoSoftware]

I actually tested them side by side on this channel already: kzbin.info/www/bejne/mqavg4F6lpJ2iJI

[@puremajik]

For a higher performance pipeline with ai, I suggest using test first development TFD pattern

[@titulusdesiderio9116]

what pattern?

[@MarvijoSoftware]

I'm assuming you mean TDD (Test Driven Development)? I actually wrote about this approach in a Reddit post: www.reddit.com/r/ChatGPTCoding/comments/1i1tkg6/hot_take_tdd_is_back_big_time/

[@billb2112]

this has been my experience. I want to love Cline, but I'm just way more productive in Cursor.

[@deadbeafc001]

Clain (and its forks) send a huge number of system prompts each turn, wasting input tokens even with a prompt cache mechanism. On the other hand, the cursor agent consumes one fast access credit out of every month's fast access credit, even for something like sending to the terminal to ‘ls’. Perhaps the world's AI product developers are the only rich people in the world, except me.

[@MarvijoSoftware]

I agree, these AI Coders are still a bit expensive. However, I think Cursor is still worth the money for the productivity gains you get. Also, DeepSeek has made tools like Cline to be bearable

[@worthstream]

Thanks for the comparison. It did not go the way I expected, tbh. I'll have to test cursor more, even if the subscription model is not appealing.

[@MarvijoSoftware]

I also didn't expect this at all! I even say it when Cline was struggling

[@DanielWeikert]

That's interesting. Can you elaborate on the "moody" thing. I was kind of irritated when one day a complete test web app creation went smoothly and another day cursor was not able to fix bugs and eventually i dropped the test project

[@MarvijoSoftware]

By "moody" I meant exactly what you just described, perfectly! It seems to have been more consistent after recent updates. I even saw it has a feature which shows the LLM's lazy coding, "//... existing code here", but still applies the diffs correctly, without removing our code, as it did before. I left Windsurf and back to Cursor because it had more unproductive moods than productive ones

[@alekai2178]

If Cline used vector databases instead of tree-sitter do you think it would become comparable to cursor? Or are there other factors at play?

[@MarvijoSoftware]

I actually think it would be comparable. If I have some free time I might use Cursor to build a Milvus vector db into Cline and test its performance. I'll let you know

[@alekai2178]

@@MarvijoSoftware That would be awesome. Maybe also storing Cline's base tools and local MCP tools in vector db could help reduce system prompt bloat, especially as these grow. Would be nice to have an external vector db for a 3rd party MCP tool directory, but using a vetted source would be needed.

[@obetishere9215]

I canceled cursor because of cline. I find cline smarter if you provide the context not the entire codebase. I just specify the files that it needs to update and investigate

[@MarvijoSoftware]

@@obetishere9215 Indeed it does shine in that context. Aider as well because they both use tree-sitter

[@tecnopadre]

Can you give more details about this? I use Roo Cline. Thanks!

[@obetishere9215]

@ i usually tell the llm what files to check and update if i need something to be implement. It makes it more reliable. Just like guiding a junior dev. My issue with cursor is it deletes some files, code that i told not to. Maybe my prompting and guidance suits more with cline.

[@RealLexable]

RooCline would also have been interesting. More than Cline alone i think.

[@MarvijoSoftware]

I'll give it a shot! Though they don't have enough differences to warrant a full test, what do you think?

[@alyax2026]

This voice is much better than others. Also, how do I use Cline computer use?

[@MarvijoSoftware]

You just tell it to. i.e., "Use Computer Use to..."

[@OG-SA]

Please provide us with the [Rules for AI] you used, as well as the CHANGLOG md NewKnowledgeBase md How did you do it ? Please I need your guidance and prompts/rules Thank you a lot.

[@MarvijoSoftware]

Hi @OG-SA. I created a repo for you with all the files you requested: github.com/marvijo-code/marvijo-software-yt Note: the files might've changed after the end of the video because the tools auto update the CHANGELOG and NewKnowledgeBase as they work. I tried retrieving them from the last commit from the video

[@OG-SA]

@@MarvijoSoftware Thank u I learned a lot from your video, keep it up

[@rashad1916]

@@MarvijoSoftwareThis is gold. This knowledge base tip is fire. Thank you!!!!! Do you include the knowledge base in the chat? Or do you make it a rules file?

[@MarvijoSoftware]

Resources repo: github.com/marvijo-code/marvijo-software-yt Cursor and Cline: CHANGELOG, NewKnowledgeBase Cursor: Instructions for AI Cline: Custom Instructions Note: CHANGELOG and NewKnowledgeBase include data from previous sessions

[@toddschavey6736]

Like Cursor, I think Cline was having a bad day. These LLMs are non-deterministic 🎲 That being said, i was pretty impressed by Cursor here. Hypothesis: You must prompt differently to Cline then you would Cursor. As long as I am talking to Cline as a equal co-worker who is both remote and suffers from short term memory...we an awesome team 😅 Oh and .clinerules 🎉

[@MarvijoSoftware]

I agree! They're non-deterministic. I even used the word "moody" in a Reddit post

[@hammeedabdo.82]

What is the best AI coder?

[@MarvijoSoftware]

I wouldn't say it's definitive currently, keep watching the channel, we'll have a definite winner soon. The thing is that the Coders weren't tested with larger codebases, so I'm filling that gap

[@rafael_tg]

Aider. But you need to be more specific in the prompts

[@titulusdesiderio9116]

From my very experience, Cursor is more powerful fore sure. But Cline with Gemini 2.0 and/or Sonnet under the hood is pretty ok. It's solves about 8 of 10 issues and tasks/features. Cursor does 9. And the last 10'th solvable too if I use o1

[@MarvijoSoftware]

o1 is definitely a good fallback when Sonnet can't figure things out. I used Cline with Sonnet in this review. To be honest I didn't expect Cline to struggle with such straightforward bugs

[@boynet2]

I have similar experience with cline, I believe the massive system prompt is making the models dumber.. because when I copy the same prompt and files that cline read into chatgpt it easily solve the tasks, its must be the system prompt

[@MarvijoSoftware]

Agreed!

[@tianyili6336]

A huge chunk of the prompt >50% is an example of a mcp server. If you don’t build mcp server, deleting that would help a lot

[@MarvijoSoftware]

Aider vs Cline Using DeepSeek 3: kzbin.info/www/bejne/m2LSdYqbjtuDmLs

[@MarvijoSoftware]

Aider + Deepseek 3 vs Claude 3.5 Sonnet: kzbin.info/www/bejne/e4a7eoata9ynq9E

[@mgan59]

Clearly the ai voice generator can’t pronounce “web scraping”