I got into Cursor a little late, went really heavy on it, and see myself using it less and less as I go back to VSCode.
1) The most useful thing about Cursor was always state management of agent edits: Being able roll back to previous states after some edits with the click of a button, or reapply changes, and preview edits, etc. But weirdly, it seems like they never recognized this differentiator, and indeed it remains a bit buggy, and some crucial things (like mass-reapply after a rollback) never got implemented.
2) Adding autocomplete to the prompt box gives me suspicion they somehow still do not understand best practices in using AI to write code. It is more crucial than ever to be clear in your mind what you want to do in a codebase, so that you can recognize when AI is deviating from that path. Giving the LLM more and earlier opportunities to create deviation is a terrible idea.
3) Claude Code was fine in CLI and has a nearly-identical extension pane now too. For the same price, I seem to get just as much usage, in addition to a Claude subscription.
I think Cursor will lose because models were never their advantage and they do not seem to really be thought leaders on LLM-driven software development.
1. Checkpoints/rollbacks are still a focus for us, albeit it's less used for those working with git. Could you share the bug you saw?
2. Autocomplete for prompts was something we were skeptical of as well, but found it really useful internally to save time completing filenames of open code files, or tabbing to automatically include a recently opened file into the context. Goal here is to save you keystrokes. It doesn't use an LLM to generate the autocomplete.
3. A lot of folks don't want to juggle three AI subscriptions for coding and have found the Cursor sub where they can use GPT, Claude, Gemini, Grok models to be a nice balance. YMMV of course!
I find the amount of credits included in the pro subscription per month totally insufficient. Maybe it lasts 1-2 weeks.
Today I got a message telling me I exhausted my subscription when the web dashboard was showing 450/500. Is there a team level constraint on top of individual ones?
Addressing 2) first: That's good, I totally misunderstood then, and guess I'll need to try it to understand what's new since I thought that kind of tabbing had been there a while.
Back to 1): The type of bug I see most often is where conversation history seems incomplete, and I have trouble rolling back to or even finding a previous point that I am certain existed.
Git shares some features but I think Git was not made for the type of conversational rapid-prototyping LLMs enable. I don't want to be making commits every edit in some kind of parallel-git-state. Cursor's rollback and branching conversations make it easy to backup if a given chat goes down the wrong path. Reapply is tedious since it has to be done one edit at a time - would be nice if you could roll-forward.
I haven't put much thought into what else would be useful, but in general the most value I get from Cursor is simplifying the complex state of branching conversations.
FWIW, my workflow with git is to stage changes I want to keep after every prompt. Then I can discard changes in the working area after a bad prompt or stage individual changes before discarding from the working area. Works really nice for me.
i really tried to use cursor and really wanted to like it but i constantly ran into this problem where the agent wasnt able to see what was happening in the terminal.
Since we have cursor people joining, let me bring up my constant problems around applying code changes. For background, I mostly work with "chat":
1. The apply button does not appear. This used to be mostly a problem with Gemini 2.5 Pro and GPT-5 but now sometimes happens with all models. Very annoying because I have to apply manually
2. Cursor doesn't recognize which file to apply changes to and just uses the currently open file. Also very annoying and impossible to change the file to which I want to apply changes after they were applied to one file.
For both of these scenarios, it seems to happen when the context limit is getting full and the context is summarized. I've found it usually works to respond with the right file, i.e. "great, let's apply those changes in @path/to/file", but it may also be a good time to return to an earlier conversation point by editing one of your previous messages. You might edit the message that got you the response with changes not linked to a specific file, including the file path in that prompt will usually get you back on track.
> 2. Autocomplete for prompts was something we were skeptical of as well, but found it really useful internally to save time completing filenames of open code files, or tabbing to automatically include a recently opened file into the context. Goal here is to save you keystrokes. It doesn't use an LLM to generate the autocomplete.
Oh ok, thanks for clarifying. That indeed seems like would be helpful.
Should still be configurable to turn off though (like any auto-complete, right)
I tried vanilla vs code again three weeks ago. The tab complete was so slow and gave such poor results that I had to crawl back to Cursor after not even a full week. Cursor is sooo fast and useful in comparison.
I enjoyed tab complete at first, but realized after a while that it was subtly changing what I meant to do. Too many suggestions forcing me to stop and read, interrupting the clarity I had on what needs done.
The main reason I use Cursor is for their tab complete model. I enjoy using Neovim more than Cursor, but I choose not to use it because I haven't been able to find anything that makes me as productive as Cursor's tab model.
Agreed. If I'm typing code, it's probably because I know what I want to do, and suggestions popping up easily break my train of thought. I also hate the choice of tab - I use tab for...tabs.
There is probably a configurable way to request suggestions instead of having them automatically pop up - I should configure that.
Specifically, alt+tab to enable/disable Cursor tab, and change hotkey for accepting suggestions to something besides tab (whoah, we can use [TAB] for coding again?).
> Adding autocomplete to the prompt box gives me suspicion they somehow still do not understand best practices in using AI to write code. It is more crucial than ever to be clear in your mind what you want to do in a codebase, so that you can recognize when AI is deviating from that path. Giving the LLM more and earlier opportunities to create deviation is a terrible idea.
Agreed 100%
Any time there's LLM auto complete on the prompt (chatgpt has done this too!) I find it horribly distracting and it often makes me completely lose track of what I had in mind, especially on more difficult tasks.
What I hate about Cursor is that - even when I have credits left - it still hangs for a full minute before starting to respond. Not always, but often enough. You don't know what you're buying, you might get fast response, you might get a tortoise.
I totally agree on point 1. Being able to make piece wise, limited updates with AI was a sweet spot, but they keep pushing towards “ai changes hundreds of lines across dozens of files” type edits. I bought into cursor heavily from the off, and I’ve seen it create auth bypasses, duplicate component libraries and break ORM models. I know what I want to do, I just want it to happen faster and in a way I can control and direct, and that’s not the direction cursor seems to to be going.
I’ve actually gone back to neovim, copying in snippets from ChatGPT. I don’t think I’ve given up anything in speed.
I do find the agents useful sometimes, but Cursor is most useful when it gives me visibility and granular control over what the agents are doing. Trying to follow a verbose narrative through a narrow chat pane is not it.
I have tried similar workflows (Neovim + Opencode/Codex CLI), and for me, the biggest downside compared to Cursor is the lack of a tab completion model as good as Cursor's. Supermaven is the best one I've found so far for Neovim, but it gives worse suggestions and can only suggest changes on the same line you are on.
> Claude Code was fine in CLI and has a nearly-identical extension pane now too.
One of us is wrong here. Last I checked, the extension pane was a command line, that doesn't use macOS keybindings, reimplements common controls, uses monospaced text for prose, etc.
I don't mind particularly about the last two but 'cmd A' on my Mac highlight all the text in the Claude Code user interface, rather than the text in the text box, is annoying.
That's exactly the one I'm discussing. The extension pane runs in a terminal and has all the behaviors mentioned. Unless they rewrote the thing in Electron it doesn't use normal UI keyboard shortcuts or controls.
VScode has a bit of a history now of quickly deprecating competitors who innovate in this space. It already has good options for code completion, AI chat bots, and more features on the horizon. I'm not sure what cursors moat is. Seems to me like Microsoft could easily implement any new feature cursor comes up with.
That is why most forks lose in the end, be it egcs, deno, NoSQL or cursor, the leaders eventually integrate the key features that have made them a differentiatior, thus people don't have to change, lose the investement they already had on the existing tool, and get those features as well.
For me, the best kind of "moat" (tbh I hate that word, since it specifically implies needing to design (...scheme...) and engineer some kind of user lock-in, which is inherently user-hostile) would be staying aggressively on the forefront of DX. More important than feature churn, making it polished and seamless and keeping a smile on my face as I work is the best kind of "moat."
It requires constant attention and vigilance, but that's better for everyone than having some kind of "moat" that lets them start coasting or worse— lets them start diverting focus to features that are relevant for their enterprise sales team but not for developers using the software.
Companies really should have to stay competitive on features and developer happiness. A moat by definition is anti-competitive.
Cursor seems to give me access to a lot of models for a single fee. I would love to just pay for Claude and maybe ChatGPT or Grok, but it seems like that's more expensive than Cursor.
Cursor was good for a little while until VSCode opened up the APIs for AI editing. Now Copilot is really good and other extensions (specifically Kilo Code) are doing things so much better!
I am seeing a lot of folks talking about maintaining a good "Agent Loop" for doing larger tasks. It seems like Kilo Code has figured it out completely for me. Using the Orchestrator mode I'm able to accomplish really big and complex tasks without having to design an agent loop or hand crafting context. It switches between modes and accomplishes the tasks. My AGENTS.md file is really minimal like "write test for changes and make small commits"
I feel like I've hit a sweet spot for my use case, but am so behind the times. I've been a developer for 20 years and I'm not interested in vibe coding or letting an agent run wild on my full code base.
Instead, I'll ask Cursor to refactor code that I know is inefficient. Abstract repetitive code into functions or includes. Recommend (but not make) changes to larger code blocks or modules to make them better. Occasionally, I'll have it author new functionality.
What I find is, Cursor's autocomplete pairs really with with the agent's context. So, even if I only ask it for suggestions and tell it to not make the change, when I start implementing those changes myself (either some or all), the shared context kicks in and autocomplete starts providing suggestions in the direction of the recommendation.
However, at any time I can change course and Cursor picks up very quickly on my new direction and the autocomplete shifts with me.
It's so powerful when I'm leading it to where I know I want to go, but having enormous amounts of training data at the ready to guide me in best-practices or common patterns.
I don't run any .md files though. I wonder what I'm missing out on.
Abstraction for abstraction sake is usually bad. What you should aim for is aligning it to the domain so that feature change requests are proportional to the work that needs to be done. Small changes, small PRs.
Did something change with Kiro, or was I just using it wrong? I tried to have it make a simple MCP server based on docs, and it seriously spent 6 hours without making a basic MVP. It looked like the most impressive planner and executor while working, but it just made a mess.
Cursor will soon be irrelevant. The one thing that it excels at is autocomplete, and that's ironically the one feature they bought out and integrated (Supermaven) instead of developing themselves, which sums it up quite well. It's still making good money based on having been first to market , but it's market share has been in freefall for ages, only accelerating.
The agentic side is nothing special and it's expensive for what you get. Even if you're the exact target audience - don't want CLI, want multiple frontier models to choose from for a fixed monthly price - Augment is both more competent and ends up cheaper.
Then for everyone else who is fine with a single model, Claude Code and now Codex are obviously better choices. Or those who want cheaper and faster through open weights models, there's Opencode and Kilo.
The mystery is that the other VC backed ones seemingly don't care or just don't put enough resources into cracking the autocomplete code, as many are still staying with Cursor purely for that - or were until CC became mainstream. Windsurf was making strides but now that's dead.
Do not underestimate enterprise customers who buy Cursor for all their employees. Cursor will become legacy tech soon, yes. But it will be slow death, not a crash.
> and that's ironically the one feature they bought out and integrated (Supermaven) instead of developing themselves
What? Cursor bought Supermaven last November and I have been using their much superior (compared to GH Copilot) completion since maybe early last year so it does not add up.
I use Cursor because I found their autocomplete to be the best option at the time. That seemed to be the consensus at one point too from bits of research I did.
Do people think there are better autocomplete options available now? Is it a case of just using a particular model for autocomplete in whatever IDE you want to use?
Right. Yesterday I tried a simple task that just adds Required[] notation to all class fields. After making the change on one field, Cursor allows me to press tabs and update all other fields. VSCode doesn't understand what I was trying to do after the first operation, which is surprisingly bad (no improvement after months). Also I'm not in favor of the conversational experience of claude code or other CLIs for such trivial task.
I'd be happy to know what else can provide a better user experience than Cursor.
Disclaimer: I get enterprise level subscriptions to these services via my employer. I personally don't pay for them and never consider their cost, if that matters.
The reason is abundantly clear. Cursor was just a GPT wrapper with a nice UI/UX (which was very nice when it came out) it has some other models like autocomplete as well, but its still a wrapper. OpenAI and Anthropic build and train models specifically to work via CLI driven processes, which is why they are so much better now. Cursor is basically dead as I'm sure they realized they get much better performance with the CLI/agentic approach.
> OpenAI and Anthropic build and train models specifically to work via CLI driven processes,
Cursor agents open terminals just fine in VSCode and is a major part of how Cursor works.
I personally code in VSCode text editor prior to Cursor (left VIM a while ago) and prefer to stay in the context of a desktop text editor. I find it's easier to see what's changing in real time, with a file list, file tabs, top level and inline undo buttons etc.
I've even stopped tabbing to a separate terminal by about 50%, I learned to use VSCode terminals to run tests and git commands, which works well once you learn the shortcuts + integrate it into some VSCode test runner extensions. Plus Cursor added LLM/autocomplete to terminal commands which is great. I don't need a separate CLI tool or Bash/zsh script inside terminal to inject terminal commands I forgot the arguments for.
you're missing that open ai and anthropic are finetuning their models for their coding agents, i.e. using rl and annotating datasets and optimizing for them with directly learnable parameters. Cursor is just able to use whatever foundational model apis are available (not sure if or when labs might give programmatic access to coding agents). I'm sure they are trying to train OSS models but those fall way short performance wise of proprietary ones.
I'd love to hear from folks who mainly use Claude Code on why they prefer it and how they compare. It seems to be the most popular option here in HN, or at least the most frequently mentioned, and I never quite got why.
I always preferred the deep IDE integration that Cursor offers. I do use AI extensively for coding, but as a tool in the toolbox, it's not always the best in every context, and I see myself often switching between vibe coding and regular coding, with various levels of hand-holding. And I do also like having access to other AI providers, I have used various Claude models quite a lot, but they are not the be-all-end-all. I often got better results with o3 and now GPT-5 Thinking, even if they are slower, it's good to be able to switch and test.
I always felt that the UX of tools like Claude Code encourage you to blindly do everything through AI, it's not as seamless to dig-in and take more control when it makes sense to do so. That being said, they are very similar now, they all constantly copy each other. I suppose for many it's just inertia as well, simply about which one they tried first and what they are subscribed to, to an extent that is the case for me too.
I don't think we are in a phase where we can confidently state that there's a correct answer on how to do development, productivity self reports are notoriously unreliable.
At least personally, the reason why I prefer CLI tools like Claude and Codex is precisely that they feel like yet another tool in my toolbox, more so than with AI integrated in the editor. As a matter of fact I dislike almost all AI integrations and Claude Code was when AI really "clicked" for me. I'd rather start a session on a fresh branch, work on something else while I wait for the task to be done, and then look at the diff with git difftool or IDE-integrated equivalent. I'd argue you have just as much control with this workflow!
A final note on the models: I'm a fan of Claude models, but I have to begrudgingly admit that gpt-5-codex high is very good. I wouldn't have subscribed just for the gpt-5 family, but Codex is worth it.
It's primarily the simplicity with which I can work on multiple things. Claude code is also very good with using tools and stuff like that in the background so I just use a browser MCP and it does stuff by itself. I hook it up to staging bigquery and it uses test data. I don't need to see all these things. I want to look at a diff, polish it up in my IDE, and then git commit. The intermediate stuff is not that interesting to me.
This suddenly reminded me that I have a Cursor subscription so I'm going to drop it.
But of course if someone says that Cursor's flow suddenly 2x'd in speed or quality, I would switch to it. I do like having the agent tool be model hotpluggable so we're not stuck on someone's model because their agent is better, but in the end CC is good at both things and codex is similar enough that I'm fine with it. But I have little loyalty here.
That makes sense. Personally I have rarely gotten truly satisfactory results with such a hands-off approach, I cannot really stop babysitting it, so facilities to run it in the background or be able to do multiple things at once are rather irrelevant to me.
But I can see how it might make sense for you. It does depend a lot on how mainstream what you are working on is, I have definitely seen it be more than capable enough to leave it do its thing for webdev with standard stacks or conventional backend coding. I tend to switch a lot between that and a bit more exotic stuff, so I need to be able to fluidly navigate the spectrum between fully manual coding and pure vibe coding.
I think personally I really like Claude, but our company has standardized on Cursor. Both are very good. I do like the tab completion. The "accept/undo" flow of cursor is really annoying for me. I get why its there, but it just seems like a secondary on top of Git. I usually get everything in a completely committed state so I can already see all my changes through the standard git management features of "VSCode".
I think Claude's latest VSCode plugin is really great, and it does make me question why Cursor decided to fork instead of make a plugin. I'd rather have it be a plugin so I don't have to wipe out my entire Python extension stack.
I like the "accept/undo" feature because it allows for much more granular control. You can accept some files or lines, and give feedback or intervene manually in other parts. I don't like building up technical debt by accepting everything by default.
As in chess, stock trading, and combat aviation, people at first
believed humans ought to curate computer-generated strategies. Then it
become obvious the humans were unnecessary.
No doubt, I am simply being pragmatic. I will keep hand-holding AI when needed, it is increasingly less needed, good. I am not a skeptic, I will keep using AI to the limits of its ability the whole way, but one quickly learns its limits when you try to do some professional work with it.
It’s still plenty useful of course, but it absolutely needs constant babysitting for now, which is fine. I like AI coding tools that acknowledge those limits and help you work around them, rather than just pretending its magic and hiding its workings from you as an autonomous background process. Maybe soon such need for control will become obsolete, awesome, I will be the first one onboard.
PS:
Chess AI is definitely superhuman now, but Stockfish is a small NN surrounded by tons of carefully human-engineered heuristics and rules. Training an LLM (or any end-to-end self-supervised model) to be superhuman at chess is still surprisingly hard. I did some serious R&D on it a while back. Maybe we’ve gotten there in the last few years, not sure, but it’s very new and still not that much better than the best players.
Most real-world stock trading is still carefully supervised and managed by human traders. Even for automated high-frequency trading, what really works is to have an army of mathematicians devising lots of trading scripts, trading with proper deep-learning/reinforcement-learning is still surprisingly niche and unsuccessful.
Also combat aviation is far from being automated, sure they can bomb but not dogfight, and most drones are remote controlled dumb puppets.
I do agree with your point generally, but any good engineer needs to understand the details of where we are at so we can make real progress.
You can just use the official Claude Code, OpenAI Codex, and Gemini extensions on VS Code. You get diffs just like in Cursor now. The performance of these models can vary wildly depending on the agent harness they're on.
The official tools won't necessarily give you the best performance, but they're a safer bet for now. This is merely anecdotal as I haven't bothered to check rigorously, but I and others online have found that GPT-5-Codex is worse in Cursor than in the official CLI/extension/web UI.
That is a bit uncalled for, I like to be lean and technically precise as much as the next guy.
I am not talking about "deep IDE integration" in a wishy-washy sense, what I care about as a professional engineer is that such an integration allows me to seamlessly intervene and control the AI when necessary, while still benefiting from its advantages when it does work well on its own.
Blindly trusting the AI while it does things in the background has rarely worked well for me, so a UX optimized for that is less useful to me, as opposed to one designed to have the AI right where I can interlieve it with normal coding seamlessly and avoid context-switching.
I could be wrong about this, but it feels like Cursor is less and less compelling with better models and better CLI tools popping up. Are the plan limits generous enough that it's worth a spin?
Again, I haven't used Cursor in a while, I'm mostly posting this hoping for Cunningham's Law to take effect :)
Cursor is your best option if you want to switch models frequently, run multiple agents in parallel, and also have the best tab complete out there. And you're still getting extra vc-funded tokens. You get ~$40 worth of tokens at API costs for the $20 plan.
idk seems worth it to me. If youre shelling out on one of the $200 plans maybe its not as worth it, but it just seems like the best all in one ai product out there.
Does anyone have autocomplete that is half as good as Cursors? I just tried vanilla vs code with github copilot and it was terrible. Not worth paying for bad.
I find Cursor at the same level as Claude code, with some strengths and some weaknesses. Cursor is nice when I want to start multiple parallel agents, while browsing files, monitoring the progress, and switching models as needed. It’s just a simple, zero config environment i can just start using intuitively.
Claude code is more reliable and generally better at using MCP for tool cal, like docs from contex7. So if I had only one prompt and it HAD to make something work, Claude code would be my bet.
Personally I like jumping between models and IDEs , if only to mix it up. And you get a reminder of different ways of doing stuff.
I tried Claude Code once and half an hour later it printed $10 cost. I thought I was using the pro subscription, not the API. This makes using CC dangerous, so I am avoiding it.
I'm currently flying, and using Cursor. I have my model set to Sonnet-4, and it keeps bugging me that my usage is going to end on 10/21, 10/19, 10/13, 10/08, after just a couple hours of VERY slow LLM usage.
I wouldn't even bother with it, but my MCP coding tool I built uses Claud Desktop and is for windows only, and my laptop is MacOS. So I'm using Cursor, and it is WAY WORSE than my most simple of MCP servers (that literally just does dotnet commands, filesystem commands, and github commands).
I think having something that is so general like cursor causes the editor to try too many things that are outside what you actually want.
I fought for 2 hours and 45 minutes while Sonnet-4 (which is what my MCP uses) kept inventing worse ways to implement OpenAI Responses using the OpenAI-dotnet library. Even switching to GPT-5 didn't help. Adding the documentation didn't help. I went to claude in my browser, pasted the documentation, and my class I wanted extended to use Responses, and it finished it in 5 minutes.
The Cursor "special-sauce" seems to be a hinderance now-days. But beggars can't be choosers, as they say.
As everyone with half a brain predicted, their pricing was never meant to last. Their "limit" (base plan) is now just $20 in API credits, at slightly higher than provider token price. Sometimes they let you go a little over, but I'm not sure if that's still true.
I wish cursor would let you see how much usage in terms of $$ you have done for your month. Its really hard to see in the dashboard the individual charges tokens, but then there is no cumulative. I haven't been able to find a way to see how much of my included usage is being used besides downloading the csv and manually summing. They just give you a very unhelpful "You will use your included credits by X date"
I suppose this is by design so you don't know how much you have left and will need to buy more credits.
Thats on demand usage. Not your plan usage. You get Y credits every month before you start using on demand usage. That $XX/$YYY is how much of your on demand usage limit you used.
"Commands now execute in a secure, sandboxed environment. If you’re on allowlist mode, non-allowlisted commands will automatically run in a sandbox with read/write access to your workspace and no internet access."
I really don't understand Cursor's 30 billion dollars valuation (half of Antropic). I use it, it's not a bad tool by any means but it's so buggy from version to version. Latest bug I had, it completely stopped keeping my zsh state in the terminal, I had to downgrade.
And honestly, I'm not that sure what secret sauce is worth 30 billion dollars? The agent loop? Others can do that? The autocomplete?
not only that but the way that openai and claude have their own foundational models/agents trained to work via CLI, which will basically always be better than just cursors gpt wrapper approach.
frontier labs do finetuning of their models for software dev using the terminal/cli driven style, annotating datasets to solve programming in this fashion, and fine tuning will almost always make for better performance. Cursor as mostly a wrapper is just using the underlying foundation models in their framework and orchestrating on top of that, as opposed to doing actual learnable objectives in training to make things better.
Anyone have good recommendations for plugins integrating things like LM Studio or Ollama into Visual Studio or Jetbrains IDEs? I'd like to do more local AI processing on code bases instead of always relying on outside providers, but a lot of these things like Copilot and Cursor seem so well integrated into the IDE.
Copilot in VSCode supports local models through Ollama as well. Not sure about Copilot in Visual Studio. That's one of the most annoying things is VS is always behind VSCode in terms of Copilot features.
How do I get to the point of choosing model API providers (and configure my own) without doing any kind of sign in with Copilot though? I'm looking for everything be offline and not affiliated with some external service at all, even if some free tier.
Is it possible to run Cursor entirely with local models? My Mac can comfortably run relatively massive models. I would experiment so much more with AI in my codebases knowing that I won't slam into a brick wall due to quotas, connection issues, etc.
You can power it with local models but you can’t use it without internet, or without sending your data to cursor, since they do a bunch of preprocessing and orchestration on the backend before handing everything off to the model
Great job ya'll! Admittedly I haven't tried Claude Code but that's because I think Cursor is the bees knees! I do agree with the people posting that the 500 requests pricing approach is kinda rough, but I can see how you're trying to figure this out as a company that doesn't own a frontier model like Anthropic. Anyway, great job! _Love_ the new Agent View. Fun to see you guys hashing out this UI/UX release by release.
Seriously, every Cursor update adds another override, their most favorite seeming to be Cmd+E. If I'm not mistaken, poor Cmd+E changed what it does already 3 or 4 times. And every time I have to remove the newly added shortcut to restore "find with selection"...
Nice to see the image files being read without having to paste them and team rules. Cursor has been extremely helpful the last few months but increasingly more expensive. I spent almost 300 last month and had a lot of frustrating experiences so now I’m transitioning to Claude code in VS code.
How can I enable new sandboxed terminal? I have the latest version and trying to send "run curl cursor", but the command is not executed, it is the same behaviour like in previous versions. I don't see option Auto-Run in sandbox.
Since finding out a Claude Code extension to run on VS Code/cursor, I use it less and less. With git and Claude Code, rolling back and forth is a breeze. Cursor is cooked, as the cool kids say now days. They need to adapt and find a moat.
Was spending a lot with cursor switching between sonnet and opus 4.1s like 1500 to $2k a month. Was doing a lot of tabs in parallel of course. Output was like 5k lines on Good day. (Lines not the best measurement) But a yard stick against feature testing and rework.
Now with gpt-5-codex and codex vs code ext .. getting through up to 20k line changes in a day again lots of parallel jobs; but codex allows for less rework.
The job of the "engineer" has changed a lot. At 5k lines I was not reviewing every detail but it was possible to skim over what had changed. At 20k it's more looking at logs performance / arch & observation of features less code is reviewed.
Maybe soon just looking at outcomes. Things are moving quickly.
Sounds like a different use case than Cursor. Editing that many files/lines probably scales better with a CLI tool. Cursor makes more sense for day-to-day maintenance and heavy hand-holding feature development coding.
If I was building a new project from scratch I'd probably use a CLI tool to manage a longer TODO easier. But working on existing legacy code I find an IDE integration is more flexible.
Like many others I was very pro cursor a year or so ago, but unfortunately since then 3 significant things have severely impacted its appeal:
VS Code accepted the challenge and upped its game.
Claude Code changed the game.
Cursor's own heavy value decrease (always part of the strategy but poorly communicated and managed) hit Cursor users hard when the cheap premium tokens honeymoon ended in recent months.
Existing users are disappointed, potential new users no longer see it as the clear class leader, because it isn't.
I use Cursor only because of their extraordinary tab suggestion feature. It's still blowing the competition away. It's 1000000 times better than any other tools I've tried and I've tried them all. Cursor should focus on that feature alone - anything else will end up in enshittification.
I used to work at NASA as an engineer, and the way it works (E.g space station operations or shuttle missions) was that hundreds of engineers are working on various complex systems on the ground while astronauts try to do a mission in space.
What this means that dozens of procedures and activities are happening at any one time in orbit, and that a Flight Director on the ground and an Astronaut in space needs to be at least cognizant or aware of (at least enough to prevent disasters and complete the tasks) -this is the greatest challenge in on-orbit work.
IE to widen this metaphor: to collect and gather complex operational data on differing parts of a system in a USEFUL way is the greatest challenge of complex work and software engineering is about controlling complexity above all.
Now at NASA, we often wrote up procedures and activities with the- "Astronauts are smart they can grok it" mindset; but during debriefs the common refrain from those at the top of pyramid was that "I don't have the mental capacity to handle and monitor dozens of systems at the same time" -Humans are very bad at getting in flow when monitoring things.. maybe if some kind of flow state was achievable like a conductor over an orchestra in orchestrating agents..but I don't see that happening with multiple parts of the codebase getting altered at the same time by a dozen agents.
Cursor and Agentic tools bring this complexity (and try to tame it through a chat window or text response) to our daily work on our desktop; now we might have dozens of AI Agents working on aspects of your codebase! Yes, its incredible progress but with this amazing technical ability comes great responsibility for the human overseer...this is the 'astronaut' in my earlier metaphor- an overburdened software engineer.
Worryingly also culture wise- management teams now expect software devs to deliver much faster, this is dangerous since we can use these tools but are forced to leave more to autopilot in hopes of catching bugs in test etc - I see that trend is to push away the human oversight into blind agents but this is the wrong model I think for now -how can I trust and agent without understanding all that it did?
To summarize, I like both Cursor and Claude Code, but I think we need better paradigms in terms of UX so that we can better handle conflicts, stupid models, reversions, better windows on what changed code-wise.. I also get the trend of creating trash-able instances in containers and killing them on failure, but we still need to understand how a code change impacts other parts of the codebase -
anyway somebody on the cursor team will not even read this post -they will just summarize the whole HN thread with AI and implement some software tickets to add another checkbox to the chat window in response.. this is not the engineering we need in response to this new paradigm of working.. we need some deep 'human' design thinking here..
Off topic, but does anyone understand why Apple’s reader mode is so bad? This post is an example of reader mode not displaying section titles. I see this pretty frequently, even in my own blog, and haven’t been able to figure out hires to beat its flawed logic.
Why waste precious milliseconds typing complete sentences to your AI coding assistant? With autocomplete in the prompt box, we've solved the most pressing problem facing developers today: prompt fatigue.
Gone are the days of exhausting yourself by typing full requests like "refactor this function to use async/await." Now, simply type "refac—" and let our AI predict that you want an AI to refactor your code.
The builders are quietly learning the tools, adopting new practices and building stuff. Everyone else is busy criticizing the tech for its shortcomings and imperfections.
> The builders are quietly learning the tools, adopting new practices and building stuff
I thought the "you're not a real programmer if you don't use AI" gatekeeping would take a little longer than this, but here we are. All from the most minor of jokes.
It's not a criticism of AI, broadly, it's commentary on a feature designed to make engineers (and increasingly non-engineers) even lazier about one of the main points of leverage in making AI useful.
Autocomplete is one of Cursor's most popular features, and is cited as the only reason some people continue to use it. And you're mocking the Cursor team for adding it to the one place where devs still type a lot of text, and making a value judgment by calling it lazy.
Anyone seriously using these tools knows that context engineering and detailed specific prompting is the way to be effective with agent coding.
Just take it to the extreme and youll see; what if you auto complete from a single word? A single character?
The system youre using is increasingly generating some random output instead of what you were either a) trying to do, or b) told to do.
Its funny because its like,
“How can we make vibe coding even worse?”
“…I know, lets just generate random code from random prompts”
There have been multiple recent posts about how to direct agents using a combination of planning step, context summary/packing, etc to craft detailed prompts that agents can effectively action on large code bases.
…or yeah, just hit tab and go make a coffee. Yolo.
This could have been a killer feature about using a research step to enhance a user prompt and turn it into a super prompt; but it isnt.
What’s wrong with autocompleting the prompt? There exists entropy even in the English language and especially in the prompts we feed to the llms. If I write something like “fix the ab..” and it autocompletes to AbstractBeanFactory based on the context, isn’t it useful?
> adding it to the one place where devs still type a lot of text
Because that's where the text the devs type still matters most.
Do I care significantly about this feature's existence, and find it an affront to humanity? No.
But, people who find themselves using auto-complete to make even their prompts for them will absolutely be disintermediated, so I think it wise to ensure people understand that by making funny jokes about it.
I'm currently training local LLMs on data derived by small movements of my body, like my eyes and blinking patterns, in order to skip the keyboard altogether and enter a state of pure vibe.
In fact, this entire response was written by an LLM trained on my controlled flatulence in order to respond to HN posts.
It's already been here for a long time actually. Think google search auto completion of prompts. You're looking for something that might have biases on either side, and you are only shown autocomplete entries for a specific bias.
you're right, i guess it's only negative if you think it's important for people to understand the code they produce, if that's not a concern for you then no problemo
This brings up an interesting point that's often missed, IMO. LLMs are one of the few things that work on many layers, such that once you have a layer that works, you can always add another abstraction layer on top. So yes, you could very well have a prompt that "builds prompts" that "builds prompts" that ... So something like "do x with best practices in mind" can turn into something pretty complex and "correct" down the line of a few prompt loops.
lowkey typing is so cumbersome though they should make an ai model that can read my thoughts and generate a prompt from them so i don't have to anymore
You are thinking too small. AI should be able to determine what my thoughts should be and execute them so I don't have to spend my precious time actually thinking.
1) The most useful thing about Cursor was always state management of agent edits: Being able roll back to previous states after some edits with the click of a button, or reapply changes, and preview edits, etc. But weirdly, it seems like they never recognized this differentiator, and indeed it remains a bit buggy, and some crucial things (like mass-reapply after a rollback) never got implemented.
2) Adding autocomplete to the prompt box gives me suspicion they somehow still do not understand best practices in using AI to write code. It is more crucial than ever to be clear in your mind what you want to do in a codebase, so that you can recognize when AI is deviating from that path. Giving the LLM more and earlier opportunities to create deviation is a terrible idea.
3) Claude Code was fine in CLI and has a nearly-identical extension pane now too. For the same price, I seem to get just as much usage, in addition to a Claude subscription.
I think Cursor will lose because models were never their advantage and they do not seem to really be thought leaders on LLM-driven software development.