- https://www.reddit.com/r/ChatGPTCoding/comments/1cft751/my_experience_with_github_copilot_vs_cursor/ - I currently just use Claude via Projects, ChatGPT, and Github Copilot You could also use something like https://softgen.ai/, where it is just AI-powered coding. ## Reflections I think you need to have good skills for narrowing down tasks and realizing when you can use this tool — it shouldn't be used throughout your entire time spent coding, and there's different levels at which you can use it. Trying to use it requires me to trust the AI, and that can be problematic because the solution it gives might not consider edge cases. When you're given a solution without the thought process and intermediate steps, and you're not part of that, it's easy to overlook minor things. I think that showed during one of the pull requests I made, where I didn't think about a side effect or unexpected behavior because I didn't go through the entire thought process. But at the same time, when something is already defined and there are existing patterns, it does a pretty decent job of autonomously completing it. Especially for apps or plugins like RooCode, they were able to iterate, even though it was a bit expensive because of the large system prompt. It was able to get to a pretty decent spot for me to just fine-tune. But then the problem is, when I fine-tune it, it's harder for me to do so. I fine-tune it, but I need to understand that it's not just a process. For me, I still feel like it's better to be part of the process if you are writing maintainable code. ## Apps ### Cursor 500 limit ### Cline - Pretty useful modes, but 10k system prompt is crazy 😭 ### Enterprise - [Github Enterprise](https://www.reddit.com/r/ChatGPTCoding/comments/1inyt2s/my_experience_with_cursor_vs_cline_after_3_months/) As a result of your Github Enterprise Subscription, you get access to a Quazi private/dedicated version of Different models including Sonnet 3.5+ as a co pilot model. I discovered today, you might get an error on some calls that says API Response Failed, if you get them a lot, In Github there is a setting there that says something about "match with public code" or something, it is blocked by default… unblock it. Save and try your call again, it seems to solve it, for some reason. Also the when the task is finished, resend another promt like "we had API streaming problems while executing the previous promt, please validate the files to check if corruption occurred. There is also a backoff slider that you can set that if it hits a rate limit it will Exponentially backoff until the rate limit clears. ## Models gemini 2.0 flash for 2m context as a pair programmer with stack exchange api qwen 2.5 coder 32b can be ran locally maybe i can ## Prompts ### Fixing bugs "Reflect on 5-7 different possible sources of the problem, distill those down to 1-2 most likely sources, and then add logs to validate your assumptions before we move onto implementing the actual code fix" ### Cline From a Reddit post: I use a tool called cline that integrates with VSCode a popular editor. I split big problems into smaller tasks and I ask the model to solve the small tasks. For each task you need to figure out the context that the model needs. It is usually files already in the project or documentation. Then you try to be crisp about what it needs to do. The tools then generate a diff, I inspected it closely. I have a rough idea of what code I expect to generate so it is simple to accept the code or tweak the prompt, usually by adding context.