Guide:How to Handle Big Projects With Cursor

I understand your frustration with Claude 3.5 struggling with large codebases. Based on the transcript you shared, I can provide some helpful advice on how to use PRD (Product Requirement Document) and RFC (Request for Comments) approaches with Cursor to manage large projects more effectively.

Using PRD and RFC Approach with Cursor

The approach demonstrated in the Turkish YouTube transcript is quite powerful for working with large codebases. Here’s how you can implement it:

1. Create a PRD Document

Start by creating a clear product requirements document that outlines:

  • What your application does
  • Required technologies
  • Design patterns
  • Technical specifications

2. Establish Cursor Rules

Create a set of rules for Cursor to follow when working with your code:

  • Coding standards
  • File/folder structure conventions
  • Naming conventions
  • Architecture principles

3. Create a Features Document

Break down your application into discrete features with clear specifications for each.

4. Generate RFCs for Individual Features

For each feature or change you want to implement, create an RFC document that includes:

  • Technical design
  • Required endpoints/functions
  • Data models
  • Database schema changes
  • Security considerations
  • Dependencies

5. Implementation One at a Time

Instead of asking Cursor to understand and modify your entire codebase at once, ask it to implement or modify just one specific RFC at a time.

Why This Works Better

The key insight from the transcript is that by breaking down your project into smaller, well-documented chunks, you give Cursor a clearer context that prevents it from getting lost in the complexity of your entire codebase. Each RFC provides focused guidance for a specific task without requiring Cursor to understand the entire system at once.

Example Prompt Pattern

Based on this RFC document:
[Insert your RFC details here]

Please implement/modify the following specific feature in my codebase:
[Describe exactly what needs to be changed]

Here are the relevant files:
[Include only the necessary code files for this specific change]

This approach guides Cursor’s work within clear boundaries, significantly reducing the chance of it hallucinating or making unwanted changes to your code.

Regarding your question about Claude 3.7 returning online, I don’t have specific information about when it will be back and working properly. I would recommend checking Anthropic’s official channels for the most up-to-date information on Claude 3.7’s availability.

##Resource

(It was Turkish language video)

If community users interested this topic i can record english video for how to make PRD and RFC documents for usign cursor ai.

5 Likes

I’m excited to see how much a 1 MM token window will improve this behavior.

1 Like

I for one would certainly be interested in learning how to write PRD and RFC documents for use with Cursor. I’m sure I could use an LLM to explain it to me, but if you make a video for it you’ll get at least 1 view from me!

This is very similar to the process I am using now, except this appears to be a little more structured and I’m interested to see if it works better than what I am doing currently.

Thanks for sharing!

2 Likes

I don’t have any evidence to base this on, but I don’t know if it will provide that much additional benefit. I’m sure it will be better, but I think it will be an incremental improvement.

I base this on the experience I have had thus far with LLMs. It seems that even if you don’t get anywhere near the context limit the performance of the LLMs seems to degrade with more context. It’s like there is a sweet spot where you give it enough context to understand the problem, but too much context confuses it even if you are nowhere near the context limit.
I’m not just basing this off Cursor either, I’ve noticed the same thing with chat interfaces like GPT, Bedrock, etc.

I think there are certain cases where larger context will unlock new possibilities, but for something like code generation I think additional context waters down the entire context dataset.

Like I said though, I’m just basing this off vibes and instinct, nothing concrete.

2 Likes

Yes, I too found that breaking up everything (from code to documentation) into separate files and using them in isolated contexts improves performance. For me, significantly, not just incrementally.

2 Likes

Sounds like good software design without all the overengineering.

This message was sent securely using Zix® |

  • |

We are currently in a cycle of incremental improvement, so it will be beneficial to those of us caught in that improvement cycle over the course of the next x months.