I’ve just started using cursor and have been impressed up until a few days ago.
Mainly using it to build some tools to make my job and customers lives a little easier, and it gets the basics done. But now pain points are appearing and I think there’s some issues that need to get addressed.
Cursor is too greedy, as in I need X changed, Cursor changes X,Y,Z breaking everything.
Cursor should have a method or prompt element which minimizes the footprint of changes, if you look at “fix in composer” the prompt is
“How can I resolve this? If you propose a fix, please make it concise.”
The same should be performed for all requests in composer.
Fix the bad code
VSCode has an API to access the problems tab, after any change I’m having to go through the problems and fix them one by one with composer, why? Cursor can do this with vscode’s diagnosticCollection API.
In one window cursor is interactive with git and asking to run commands
In another, it’s telling me it’s an AI that can’t execute commands and giving me a shell script.
I want cursor to have the ability to run some commands autonomously
git add / commit / diff / log
npm run build / start / test / lint
poetry / pip etc…
Obviously safe guarded, I’d even be happy to have cursor sudo’d to a less privileged account so it can just work in the background, but I want it to develop test, run them, lint the code ask me to verify that it works and then commit the code. Give me a workflow.
For cursor to really work, it needs to be features driven not code driven.
Yes we’re coders, but i just want stuff that works.
I have to do the testing, run the code and say if it work, so why not just give me a checklist to say yes this is good leave it alone now. See 1 above.
Let me tell you what’s broken and for you to fix that and only that
Build tests around it and use a workflow from #3 to ensure it stays working.
Don’t just throw random code at me
From all the tools I get to play with Cursor is close, but relying too much on the LLM.
To some extent, I agree with all you said, and I’m going to pile on a bit as there are a few things I want to add…
But there is also things you want that work, so that first :
in the composer or the chat, if you ask it to create commands and output as bash, you will have a “run” button in lieu of the “apply” one
in the composer, yesterday (first time I used it) it ran docker … commands, so I guess it can run other commands too (always ask me to confirm though, which is great !)
while the (previous docker) stack was not starting, it was trying to prompt itself to fix the problem, not succeeding as it got confused by 2 errors, impressive nonetheless ! So, to some extent, it can correct itself, by itself… feature’s probably on the way :), I sure like when it tries
now, my christmas whishlist
could I choose to use “slow” requests ? Sometimes, I don’t care about the answer coming at lightspeed, especially in the composer, while it’s trying to prompt itself.
your quota in cursor itself ? Sometimes you get carried away, a quick way of checking how much away could be nice…
Renewall date for your quota ? Monthly anniverssary, i guess, but it would be a nice too have, too. (yes…I burnt through my requests…how did you guess…)
markdown with codeblock that breaks the markdown itself… which is a shame, that’s a great way to improve your "readme"s and a nightmare to proompt you way around
bash codeblock with only a “run” button… I know…, I know, nobody codes in bash…except me probably, and cursor is great for that, but I have to ask the model for untyped codeblock to be able to apply to my files
running in circles on failing tasks in composer “surprisingly” burns through a lot of requests, it’s kind of normal, I should have kept a closer eye on proposed solutions (man, this thing makes you lazy … XD), would it be possible to “auto add” the none working solution to notepad (or suggest to add it) to mitigate that ?
same for repeated instruction from the user ? I often have to repeat instructions to ensure proper generation (yup, the markdown codeblock shenanigans, as an example)
“continue in another chat/composer” As many, I noticed that a fresh start goes a long way to fix a problem, but, you have to write a “continue” prompt, not forget stuff, bla bla bla… could we have a “one button” ((or command, or whatever ) access ?
Or/as part of, a “recurring prompt” thing, you know… how many times do you kind of repeat prompt when working ? I do, a lot… prompt snippets maybe ? so it could even be shared !
rules for file types, rules “library” (you’d use in the chat with @rule…, or something) to fine tune your prompts, quickly. I don’t want rules for my bash scripts polluting my nextjs app or the arduino code (yes, same project…) or reversly. this could be used for the “repeated instructions/tried that didn’t work” thing
chatlogs, so I (… of course I…) can go through my chat history at a later time to improve my proompt ingenearing. I would love to save a chat to a file (auto with a setting ?).
use more of git, like extract work methodology from history, commit message generation style, suggest branching/syncing/merging/… (or other workflow, configurable ? le bouchon est déjà tombé, ca coute pas bien chère de le pousser plus loin ^^) when working with chat/composer
I guess that is already a lot !
Definitly the most enjoyable AI tool I used so far ! keep up the great work
(always ask me to confirm though, which is great !)
