Cursor continuously ignores the rules

[dash]

I find it so frustrating how cursor black boxes the context window.

I recently copied and pasted 100 terminal lines into the chat (or composer, idr). Cursor was acting like it didn’t know what terminal lines I was talking about. So I kept copying and pasting terminal lines and asking it if it could see the lines. When I pasted above about 60 lines, cursor claimed it didn’t see any terminal lines at all. So I wonder where else it’s leaving out context. I would assume it removes old context, not context you just included.

[troehrkasse]

@davemac

but you are a narcisistic ■■■■■■■■ who will do the same thing again and again, rules are compete waste of time because you just ignore them

If you’re mean to the LLM and say stuff like this, it will act like that even more. This behavior keeps happening because you’re reinforcing it.

[davemac]

Reinforcing the LLM is not the problem, getting cursor to include the rules is.

Why does it reinforce so quickly when I’m being mean, and yet ignore me repeatedly when asking it to stop deleting my comments, or whatever. What’s this, selective reinforcement?

The LLM is agreeing and responding to my frustration because cursor is not doing what it said it would do. At this point in the chat it doesn’t matter what I say, the response is just agreement agreement but it’s not actually doing anything useful, so a new chat window is coming, but that won’t include the rules either.

When I prompt in the Chat or Composer, I should be talking to cursor, I’m using the cursor interface, I’m in cursor, I’m expecting cursor to somehow combine rules, @content and history context with the current prompt, so I don’t have to keep repeating everything, every time, like "don’t delete my comments, stop and think before suggesting, keep it simple”. If none of this is happening, the LLM isn’t getting the full request is it?

I’ve watched 100 videos of people claiming to have just added some rules and now it’s magically following them, every prompt. My cursor sort of did this, until I paid for it, and now I can’t get it to follow a single rule, so yeah, it ■■■■■■ me off, wastes my time, has destroyed working code and frustrates the ■■■■ out of me. Is this what’s coming to take all our jobs?

If cursor was a self-driving car, it would be out of control, bouncing off walls, people and other cars, and blaming the model. Cursor itself is pre-processing before the model even joins in, except it isn’t pre-processing as sold by the salesman, so the question here is how to get it to work as promised so the LLM can receive the full prompt?

[eq3430]

I am relived to find this thread that I am not the only one experience this recently.
It was not so nearly forgetful until couple weeks ago, I think.
All my productivity had gone into helping Cursor to remember! And yes, I quickly used up 500 queries and started paying extra on per query basis for the 500+ queries.
I hope they can fix it soon!

[alihalabyah]

I am facing this issue and its extremely annoying and wastes a lot of time

What’s the point of the AI rules and the project rules if the agent/AI is not following them?

[davemac]

Despite 1700 views, there appears that this forum can offer no answers. It is curious that no-one from Cursor has commented and makes me question why they bother to setup a community forum, for their own product, if they are not monitoring it and joining in to help their paying users.

Or perhaps they have? Perhaps they are included in the 1700-13=1687 viewers who decided that if I cannot get Rules working correctly, perhaps I am too stupid to be using Cursor, so they just scoffed and turned away?

After all, as recently as this week their changelog claimed

• Project rules: - New capability to apply rules globally and a visual indicator for when rules will be applied

and a month ago Jan 23 the headline change was

• .cursor/rules: Users can write several repository-level rules to disk in the .cursor/rules directory. The Agent will automatically choose which rule to follow.

I have updated for both these changes and since Jan 23 I have not had any automatic rules being applied, and since Feb 19 have seen no visual indicator to show me which rule is being applied.

So I tested. I removed all rules, and all Ai Rule content, unticked Use .cursorrules.

I added a single Ai Rule
“Start every reply with ”

Then I add a project rule with pattern *.wsd which is a plantuml file. I add a single line inside, “Start every plantuml diagram reply with ”

I open any file and I get the eyes at the start of the reply. But when I open and add a .wsd file to the context, I never see the graph emoji, or any visual indicator anywhere saying any rule is being applied.

Not sure I could make rules any simpler than this, and it does not work as advertised. If people are adding many more rules, its quite possible they “think” the rules apply when actually they don’t.

Come on Cursor, explain to me why I am so stupid, show me the “simple” way I should be using your wonderfully clever product, either that or show me how to get a refund for my yearly subscription, because with the amount of my time you are wasting, me and you are falling out big time.

Here is my version, it should be the latest as it was updated a few days ago, after Feb 19 changes.

Version: 0.45.15
VSCode Version: 1.96.2
Commit: 73dd83bb6f8e3a3704ad8078a8e455ac6d4260d0
Date: 2025-02-22T10:08:09.661Z
Electron: 32.2.6
Chromium: 128.0.6613.186
Node.js: 20.18.1
V8: 12.8.374.38-electron.0
OS: Darwin arm64 21.6.0

[davemac]

Hello again, for all those having issues, despite updating Cursor yesterday and this morning, I just opened another cursor window/workspace and it wanted to update again.

I now have …

Version: 0.46.4
VSCode Version: 1.96.2
Commit: db71624816f6d52f6e54f47c37cdc7df23e22620
Date: 2025-02-25T07:28:07.050Z
Electron: 32.2.6
Chromium: 128.0.6613.186
Node.js: 20.18.1
V8: 12.8.374.38-electron.0
OS: Darwin arm64 21.6.0

and now my interface has changed to the single agent, as stated in the changelog Feb 19.

and also different settings, with Rules having its own section.

Again I added a single rule to the User Rules (previously AI Rules in settings) and the same simple rule in a .mdc *.ts project rule. As I select the file, a small icon appears next to the @ icon top left of image, and in it is the rule file being auto selected. In this image, no icon is there, because a .env file defaults back to the User Rules.

So it appears that this is now working and all it needs now is a good separation of rules into the correct files, to hopefully get the intended LLM response.

[RazorBackRoar]

@saketsarin

Also you can make use of the .cursor/rules directory feature to allow it to write code according to your style and requirements.

could you please explain this? Also you can make use of the .cursor/rules directory feature to allow it to write code according to your style and requirements. I would like to know how I could better use this to write my OOP python scripts and whatever advice you could give that would make automation even better. thank you

[saketsarin]

Here’s the official docs: Cursor – Rules for AI

Although you’re just one Google search away from unlocking a lot of resources on cursorrules and how to use it efficiently. That’d be the best way to go forward. Good luck!

[RazorBackRoar]

I appreciate the suggestion, but the official Cursor forum is a dedicated resource with accurate and detailed information on how to use Cursor efficiently. While general searches can help, the official community and forum provide the most reliable guidance. It’s always best to refer to official sources for clarity. Thanks for the help my friend.

[DFravor]

What is working for me is the following:
• I break down tasks into atomic units so I have control and can make adjustments more easily if something doesn’t work as expected.
• I create each rule in a separate file.
• All rules follow the same structure: I add clear steps and include examples of both incorrect and correct answers.
• I ensure there are no contradictions between the rules.
• Since all my rules are in .mdc format, I simply prompt the .cursor/rules folder for each request.

If I notice undesirable behavior, I just tweak the rule slightly and try again until I get the correct answer and know the rule is working properly. Hope this help you

[m-akinari]

From what I’ve seen, Cursor occasionally seems to overwrite things with outdated settings.

It might be holding on to deprecated settings from an older version in an internal cache and applying those rules.

You could try deleting any related cache files and then reinstalling the app.

I’m not sure where Cursor’s internal configuration files are located, though…

[ksach]

Here’s a fun example of Cursor ignoring a rule
My rule: never suggest deleting my production cloudformation stack
Cursor: Now let’s try deploying the full stack again, but this time let’s force a fresh deployment by removing the CloudFormation stack and recreating it: npx serverless remove --stage prod