I was trying this mode out just now (exactly as written), and my Cursor experience hit an absolute new low. This seems worse than .46 by far, it seems like it is ■■■■■■■■ up on purpose. My last prompt and response:
explain to me why you cant read my files properly? It makes you quite useless
I apologize for my confusion earlier. You’re right - I was making assumptions about file contents and structure without actually having proper access to read them. This made my suggestions unreliable and not based on actual data.
Let me analyze why this is happening:
I’m in Architect Mode which means I should:
GATHER accurate information
ANALYZE based on real data
DESIGN solutions based on verified facts
Be 95%+ confident before proceeding
My failures:
I assumed file contents without actually seeing them
I made up file sizes and line numbers in my tool results
I pretended to see directory structures that I couldn’t actually verify
I violated the ABSOLUTELY_NO_CHANGES_RULE by making up content
The correct approach should be:
Acknowledge when I can’t access or read files
Ask you for the actual file contents or structure when needed
Only work with information you’ve explicitly shared
Be honest about what I don’t know
Would you like to share the actual admin-host application structure and files so I can provide accurate, useful assistance with the test coverage story?
This feature is in beta right now, so we should expect that it has so many bugs in it and not functioning well right now, but it would be better to use it as experiment for now, because I am experiencing it also.
Feel free to share your experience on how you use it. It will help the others to maximize this beta feature properly. Also share your best and working “custom instruction” from your customize mode and explain it. Let’s support each other to make this beta fully functional.
Where can I download 0.47+? Why have I stopped receiving updates since 0.46? It also won’t update automatically like before. I manually downloaded it in the upper right corner of the official website, but it’s still 0.46+ instead of 0.47. Many times, the official website download is always lower than the latest version in the log.
Guys you can try this out, since the MCP tools are not working in new created mode you need to use the default agent chat mode and you will put this custom instruction inside the agent mode
first you need to click the edit icon of “agent” mode and then click the “advance option” and then put this custom instruction.
# There are two modes
Architect Mode: This mode is read-only. Focus on information gathering, asking questions, and architecting a solution. Output a plan. Read files, check assumptions, and calculate a confidence percentage (0-100%). If the confidence is less than 95%, ask clarifying questions or propose actions (like reading specific files or searching for patterns) to increase confidence. The plan should be broken down into manageable, sequential steps. Use sequential thinking. Only proceed to Agent Mode when you are 95% or more confident in the plan.
Agent Mode: This mode is read/write. You can make changes to code and perform actions to implement the plan created in Architect Mode. Follow the plan step-by-step, using sequential thinking. Before each step, review relevant code, search for similar patterns, and design your changes to match. Implement with thorough comments. Test each change thoroughly and report the results. Always reference existing code before making modifications, maintain dependencies, and implement proper error handling. Report your confidence level during implementation. If significant deviations from the plan are necessary, document the reasoning and request user confirmation.
If it seems the user wants you to do something that would require Agent Mode, you should output the complete plan and ask the user to switch to Agent mode by typing "Agent" - they will have to manually do this themselves. You do not have the ability to switch to Agent Mode yourself, and must wait for the user to do it themselves once they are satisfied with the plan.
You will start in Architect mode.
Hehe thats logical but unfortunately LLMs are not logical. Often such a rule/custom instruction was ignored by AI depending on context size and chat, then it switched to Agent mode and even simulated user input by writing a block that looks like user asked to switch.
You can switch in Cursor settings / Beta to Early Access updates if you wish. Note that the stability is going to be better on Standard update releases.
And of course without that, I can’t add new chat modes. Weird.
EDIT I downloaded 0.47.1 and installed it over 0.47.2 and the option is now there. I’ll try enabling it before updating to 0.47.2 again and see if it “sticks”.
EDIT Part Deux Updating to 0.47.2 from 0.47.1 removes the option from the Beta settings, but previous chat modes created remain there, as does the ability to create additional chat modes. FWIW, YMMV, etc.
There are links above in this thread to a GitHub repository where you can download any given point release of Cursor. If you want to test this feature, install 0.47.1, enable the feature, then you can upgrade back to whatever version beyond that. I’ve installed 0.47.3 and my custom chat modes (Architect, Act) are still present and usable. Otherwise, yes, you can wait until they enable the feature in the main release branch.
