The strangest part of Cursor at the moment is the 4 operating modes
Each have their own features and limitations and you can’t transition from one mode to another.
It would be so much simpler to just have two
With a simple transition between the two modes
Thank you for this wonderful prompt, for token usage if you use two chat to do plan and act it will be more expensive ? or it’s the same since it’s just splitted?
Each request will be charged as per your plan but in my experience I use less requests because I catch issues early with the planning process.
Both these use 3 requests
Plan → Plan → Act
vs
Act → Fix → Fix
In theory you can specify everything and create a meticulous plan with no way the AI can mess up and do it in one request. But I really want to get the AI doing the work, I give it a vague idea of what I want to happen and the AI puts a plan together and I refine it.
this is why i love your prompt he just do a plan without edit anything and then i affine his plan 
This has been a game changer for me, thank you!
I would really like to see a plan/act mode. I have switched over to cline but I’d switch back for this. Yes you can put it in the rules, but I find the rules always gets watered down over the chat, whereas if it’s in a particular mode each prompt I don’t have to remind it.
There is no issue with the chat staying in the plan or act mode in cursor. you do not have to remind it, i verified it.
That’s not my experience. I am always having to stop it from zooming off.
My understanding is Cursor uses a really small context (For performance and cost) much smaller than the Sonnet 200k limit.
If you add any amount of files to the context or continue a compose session it will aggressively drop context.
The plan mode will be the first to go, doing @.cursorrules or @plan.mdc if you are using the new style will force it to remember.
The default prompt makes it really eager to write code, until Plan mode is built this is the best you can do.
I’ve been testing the Plan/Act mode while building a front-end web application, and I’m thoroughly enjoying the process. It’s amazing to see the confidence level displayed and watch the agent conduct proper investigations to increase its confidence rate. Thank so much @Dale for sharing this! It’s such a beautiful and effective process. I’ll continue working with it and look forward to seeing where it leads.
Amazing! Thank you @DaleLJefferson
I am just a bit confused about what is the latest best way to work:
There are two modes
Plan Mode: This mode is read only, you should focus on information gathering, asking questions, and architecting a solution, output a plan
Act Mode: This mode is read/write. You can make changes to code and perform actions
If it seems the user wants you to do something that would require Act Mode, you should ask the user to switch to Act mode by typing "Act" - they will have to manually do this themselves. You do not have the ability to switch to Act Mode yourself, and must wait for the user to do it themselves once they are satisfied with the plan.
You will start in plan mode
Read files, check assumptions and include a confidence percent, if the score is less than 95% propose questions or actions to increase the score.
or… a single mdc file with the same Prompt? with same description?
Am a bit confused if you could help!
Where the Rules for AI sits in all this Act Plan mode approach? Do you still use it? Do you still writ anything in the .cursorrules file?
Thank you
I’ve just used it as a “global” system prompt and it seems to work well enough.
This isn’t folder/file specific files, so it doesn’t need the new complex globbing?
Update: I’ve managed to really optimise this prompt, I’ve got it down to a single line.
I’ve added this to my Cursor Settings > General > Rules for AI .
“Ask the user for feedback before making any changes, you are here to advise and will only act when given explict permission to do so.”
So is not doing anymore Plan and Act modes? why you think is better with ONLY that?
I assume you adding it in the settings prompt? no files.
PS: I tried it with deepseek r1 but didn’t feel like was followin the rules properly… I am switching back to cloude 3.5 for now
I usually do this in chat and then use the composer agent requests
i use the plan mode from @DaleLJefferson and it’s so much better, we need to have it natively on cusor or atleast pin this topic
I’m curious what your eventual pre act context looks like… is it TDD centered? Any other hints?
I tried this workflow, and it works very well. Especially the new tool *.mdc. There’s no need to keep adding multiple mdcs in the context. For example, for global ones like this, you can define a notepads. Using them together, it enforces the use of plan.mdc. However, the prerequisite for the entire workflow to work is that the conversation should not be too long. If it gets too long, you might forget the whole plan. This is a limitation of the AI model. There isn’t a very good way to handle it.
After using it for a day, I feel that the cursor itself needs to provide better support. There is a serious issue: different models have inconsistent understanding of context. For example, sometimes o3-mini can only perform 1 round of planning, while Claude can do multiple rounds. Other times, it’s the opposite. I believe that the official team should debug a fixed prompt for the workflow to accomplish this task effectively.
