Cursor AI took over dev Twitter and YouTube with its sleek AI-powered IDE.
But in 2025, the backlash is real, from performance issues and price com...
For further actions, you may consider blocking this person and/or reporting abuse
I'll chime in here, I suppose. I've been using Cursor for a while and have a love/hate relationship with it. For simple or boilerplate code, it's amazing. For explaining code it works pretty well (though it sometimes gets confused if you are building something that has a name similar to a commonly used package - I have a tool I'm building for a toolkit called "Apollo" and I frequently need to remind the LLM that I am Not in fact using Apollo client or graphql).
Bug fixing is mostly negative. Sometimes it will catch things, especially if it's a typo or accidental infinite loop, but building more "complex" things, say a sortable nested folder tree with drag and drop support, it suggests poorly written code 9/10. In most cases, however, the quality of the code depends on the quality of the prompt. More advanced features must often be broken down into smaller chunks, and then you have to play the game where you connect the pieces. Which is good, in a way, because I never want to rely on LLM to my detriment. I'll also note, in case it's relevant, I only use "ask" mode - I want to evaluate the code mentally and then implement it to fit with other systems. I only use agentic mode for boilerplate or rapid scaffolding.
Performance is a pain. My laptop has 32 gb ram, and Cursor eats it up like it's at a buffet. I often need to bounce between 3-4 projects/code bases and it's easily on par with my browser (2 windows, about 20 tabs each). If I want to game after work, i often just resort to restarting my pc (Cursor is running in WSL).
Overall, I enjoy Cursor and it has helped with some complex tasks, but I do often find myself wishing it was leaner and that the LLMs worked a bit better at "deduction" (Not a Cursor issue, just Gen AI in general - helped by providing better prompts). If I have a problem that is common in stackoverflow, for example, Cursor will often get me a good solution on its first try. Anything more complex, or "senior" level code, not so much.
I need to learn more about the MCP thing though. Any advice? Only tool I have listed in it is Console Ninja.
I do agree with the general sentiment that prompt quality is directly related to output quality. I'm often reminded of the "make a sandwich" game that some people play when learning to think like a programmer. One person provides directions and the other will do EXACTLY what they are told, and the instructor has to learn to provide clear, simple step by step instructions in order to get the desired result).
Thats a sign of a product that, to me has serious design points. I predict they won't go to far given the massive MSFT resources.
I use vscode every day. Their copilot plug-in has some similar issues. Super slow response , frequent hallucinating, stubbornessstubbornness, continuation of its own design, won't pick up manual changes on suggestions. There are lots of gaps making this a mostly bad experience due to two week iterations.
It's a major impediment to delivery now. The potential keeps us trying.
Hey, really appreciate you taking the time to share such a detailed and balanced perspective.
Totally hear you on the love/hate dynamic. Cursor definitely shines when it comes to boilerplate and simpler logic. But yes, when it’s faced with nested or more “senior-level” patterns (like your sortable folder tree example), it can feel like you’re stitching half-baked puzzle pieces together from multiple prompts. Kind of fun... kind of exhausting.
That Apollo confusion sounds all too familiar, I've run into something similar when building tools with generic names, and the LLM insists on dragging in irrelevant package context.
You raised a great point about “ask mode” vs “agentic mode” too. I also prefer reviewing code mentally before integrating, almost like pair programming with a hyperactive intern. Fast, but sometimes a bit too eager.
In terms of performance, I feel you. Cursor in WSL is a memory monster. I ended up limiting its background context size and turning off certain telemetry-based features to make it bearable. Still not ideal if you’re switching between work and gaming on the same machine.
About MCP (Most Capable Programmer), it’s a feature Cursor uses to let you “declare” which tools or agents you trust most in a workspace, like Console Ninja. If you haven’t already, try adding tools like AskTheCode, Inline Suggestions, or Trace when they’re available, they expand what Cursor can do in specific coding scenarios. I might put together a follow-up post diving into MCP setups and some underrated tools inside Cursor.
Thanks again lad for jumping in. Curious to hear more about how you’re using Cursor across multiple projects, especially if you’ve found any tricks to streamline memory usage or improve LLM grounding.
Out of curiosity
By the time you have created prompt that will provide desired code won't you manage to write it?
And eventually you will start structuring natural language so it ends up just another coding language but with so many more interpretations😅
I do love generating boilerplate but filling in code myself is not slowing me down.
It helps me think what the code should actually do and what edges I haven't thought about
I find it's very fast at producing prototype level code; much faster than a senior dev even. But once that code gets moderately complex, the oversight and rewriting ends up taking as long a s doing it by hand would have.
I’ve had that exact thought myself more than once.
Sometimes writing the prompt does feel like it takes just as much brainpower as writing the code directly. Especially when you're already clear on what you want the function or logic to do, it can be quicker to just get on with it.
I've used Cursor, Windsurf, Jetbrains AI (in Pycharm and Webstorm), GH co-pilot (paid) in VScode and Pycharm.
My workflow jumps from using Claude Desktop and Claude in Cursor. But with each passing day, I'm getting serious thoughts on learning NeoVim and just using Claude to generate smaller functions and use them as I want them to.
Note: All of my workflows are custom and complex, I've have had higher success with using Claude Sonnet via Desktop/Chat instead of cursor which keeps breaking things.
Sounds like you’ve really run the gauntlet with AI tooling, love how you’ve tested across environments and zeroed in on what actually works for your workflow.
Totally get the Claude > Cursor preference too. When you’re working with custom, complex setups, stability and control matter more than shiny integrations. I’ve had similar moments where Cursor’s context injection tries to be helpful and ends up breaking things that were already working fine.
NeoVim + Claude sounds like a clean, focused stack, especially if you’re comfortable managing your own flow. Curious though, what’s pushing you toward NeoVim now? Speed? Customisation? Just needing more precision?
AI was never good for coding. Not saying it will never be good, but I am yet to see a tool that can really produce something useful beyond anything simple.
My pro-tip for ai assistants is think about it like you just hired this guy a week ago, straight out of college. He has technical know-how, but lacks any experience. I'm not sure we can even teach LLMs experience, so they will never meet even moderate expectations of a real assistant.
That’s honestly one of the best mental models I’ve heard “fresh grad with solid knowledge, zero experience” is exactly how it feels using these tools.
They can quote the textbook, but ask them to apply it in a messy real-world context with vague requirements, and they get overwhelmed or overly confident with the wrong answer.
I HATE rewriting and rewriting the code the AI produces. I DO like that it gives ideas. But, it just is more debugging to do. Too much!
I've been using windsurf since day one and it's assistant before that. While I have had a few problems with it, it is far better a fixing bugs and generating the correct boilerplate for me. Most assistants (I've tried them all) can do some of what it can do, but not all of it. My biggest beef is that it hallucinates and requires a new context (chat) much sooner than others.
I've heard really solid things about Windsurf, especially from folks who've stuck with it since the early days. Interesting to hear it beats others in bug fixing and boilerplate, that’s usually where most assistants start to fall apart for me.
I’m curious though: have you found any good workarounds for the context resets? Like chunking tasks differently or structuring prompts in a certain way?
I might give Windsurf another spin after hearing this
To minimize error you have to evolve your workflows, can't keep doing the same thing and expect different results, if you're not using the MCP servers you're operating at 50% capacity in terms of error rates and expected output, then there's ROO Cline for VSCODE, probably the most comprehensive Agentic AI I've seen thus far.
You also must upgrade your prompt game, if you're inputting simple prompts expect simple output, the more detailed you can make your prompts optimizing it for the LLM you're using you will be blown away by the results and not continue to scratch the surface with basic prompts expecting amazing results.
Leverage different AI platforms to produce amazing outputs, optimize a prompt on Gpt4o for Claude.
You can make some really amazing things with a well detailed lengthy prompt for whatever it is you're trying to build. Stop scratching the surface with basic prompts
I use Windsurf and I pay $60/mo out of pocket because my work won't pay for it. But it has made me love my job again. It is well worth the price and it is incredible with Claude 3.7.
when a tool actually makes you enjoy your job again, it's clearly doing something right. Claude 3.7’s reasoning capabilities are next level, so I can imagine Windsurf + Claude is a strong combo.
Also respect the fact you’re paying out of pocket. That’s the kind of endorsement you can’t fake.
Out of curiosity, are there any specific workflows or tasks where it’s made the biggest difference for you?
The stuff I've been dreading for months or years. Backlogged Jira tickets that I've felt we are understaffed for, refactoring temperamental parts of our products, updating documentation, documenting poorly designed legacy code, creating proofs of concepts when I don't have figma designs, and last but definitely not least- debugging problems that would have previously taken me hours to solve in minutes.
Everything about it has been a game changer.
I've been doing this work for many years now and my company has been chronically understaffed so Windsurf has given me breathing room. Plus, I already think of myself as a competent engineer so I like to think I'm using it responsibly and not writing future technical debt code.
You should definitely write a case study on it, when you've time. It would help everyone a lot to understand windsurf and how to use it productively. The way you summarised it, makes me realise that, damn you nailed it....
Honestly I just hate the agent mode. It's more trouble than it's worth usually
While these are all valid points, the main reason for me was VS Codes's insanely frequent updates. This means Cursor is constantly running behind in updates.
My nightmare moment was when I asked it a git question and in addition to a mediocre answer, it decided to keep going and run
git reset --hard origin/mainLOL...Definitely bookmarking this under "reasons I don’t give full control to anything with an autocomplete."
Why does this article feels like an AI summary?