Support for editing the prompts sent to the GPT by Cursor

[joe8meng]

Where could this project set up the ChatGPT prompts?

[truell20]

Thanks for writing in! Could you say a bit more? Are you looking to control prompt we send to the AI?

[joe8meng]

Thanks for writing in! Could you say a bit more? Are you looking to control prompt we send to the AI?

Thanks for your prompt response. Yes, you’re right. I am interested in understanding how to control the prompts sent to the AI in the context of this project. Specifically, I’d like to know how to structure the prompts for effective interaction and where in the code these prompts are defined and used.

Any guidance or documentation you could point me towards would be greatly appreciated.

[truell20]

Curious to understand why you'd want to change the prompts sent to the AI? Could you say a bit more about what prompts you'd want to send to the AI?

Apologies for the late reply here! This is something we'd be interested in supporting.

[copilotmeng]

Hey Michael, I appreciate you getting back to me.

The primary reason I'd like to customize the prompts sent to the AI is to better tailor the responses for my particular application. By modifying the prompts, I believe I can guide the AI towards producing outputs that are more in line with the specific context and needs of my use case.

[jordaaash]

I have the same question. I want to try things like

• make it answer more succinctly with fewer filler words
• skip suggesting more simplistic answers unless there's a high probability they are universally correct (don't suggest something that will be obvious to an experienced engineer)
• be willing to say faster if doesn't know something rather than making it up
• don't provide code examples unless asked to
• not apologize for errors or otherwise provide information that isn't relevant to the code, questions, and answers

To an extent, it's possible to enter these things into a given chat session, but I want to be able to customize the prompt and learn what works and have it be automated in future sessions.

[tkeith]

I would also love to see a way to introspect/modify the prompts that Cursor uses! I love Cursor and use it every day, but often hit failure scenarios like the attached screenshot. It's hard to say whether these are completely unsolvable model issues or whether the prompt being sent is contributing to a failure like this. By having access to the prompts, I could play around with different modifications and see if I'm able to incrementally improve it in a way that avoids failure scenarios like this.

Example failure scenario screenshot: gpt-4-turbo-preview getting confused between a regex and a string

[jralden]

Claude seems to periodically forget that it is running in Cursor and has access to my files. I would like to remind it in my system prompt. It will fairly frequently start telling me to insert code snippets into a file we've been working on and I have to tell it "you do it". It will then apologize and tell me it now realizes it has access to my project files.