because when i am building code with cursor and complete a functionality and move to next it breaks the previous functionality, i have to do it over and over again to fix the previously build functionalities
okay thanks i will but what about my credits, i am really struggling to work like this. it takes me 20 requests just to make a change and then another 30 requests to fix previously created functionality
We were all there starting with Cursor, burning all our fast requests in days to fix error loops or [insert dumb AI choice], the trick is to understand context is king, including rules and documentation, secondly but of utmost importance, learn to work with the model, as an example Claude thinking is hard to debug as it doesn’t go in depth about its thinking but you can ask for a Tree of Thoughts reasoning and you’ll be able to understand why its behaving in a certain way, you’ll find a treasure of tips and tricks in this forum, be curious and practice to become better at prompting and managing context.
Hi, I too when I was still on another smart IDE burned all the queries quickly, in order to keep all your smart queries from flying away in 2 days like you did, I advise you to use free chat models such as gemini 2.0 flash exp and DeepSeek R1/V3. Both of these models are almost as good as Claude 3.5. Yes, there are tasks that need to be solved in Agent mode, and tasks that only claude 3.5 solves, but most tasks are solved by gemini 2.0 and DeepSeek R1. For clarity I’m attaching a screenshot of my traffic usage, keep in mind that this doesn’t include DeepSeek model usage, so I think I used >1000 free fast queries. Also, it would be good for you to learn how to pass context to normal chat, once you learn how to do that (prompting) you’ll be surprised that normal chat does a pretty good job as well. There is also a regular composer, yes it is worse than agent, but it can handle some tasks too, and regular composer supports all free models. I put all routine on free models, claude 3.5 I use for really complex tasks where you need to analyze a lot, and even something new to invent, for example, to rewrite someone’s large library for your own tasks. This screenshot is from 15 days of work
Remember Cursor (unless you are using their model) is the UI for AI to interact with you and your code.
They cant control the AI’s response, and honeslty they cant control the quality of your prompts.
If there are bugs, file them - if not consider being more consise with your code.
Weigh it up against paying a dev team to do the same, Cursor is way better.
I tried continue, cody, roo code, github copilot, LM Studio (to run a model locally) but none interact with the code as well.
I agree, Cursor is really good. i panicked when i discovered that i lost of my credits. to be honest i didnt even know that there is a limit. but yes its really good on what it can do and also now that i know a little better my prompts are getting better too. its just a shame that i lost so much so quickly