Just working on a 1300 row frontend file + 200 row backend file and nothing else is a headache because not all of the frontend file is visible to the LLM, so it cannot fix the issues that I’m telling it about. When I’m trying to fix UX bugs regarding some situations that can occur and should be displayed in a certain way, I end up spending hours doing prompt judo and eventually just scrapping all of it.
These hours or at least half hours of prompt judo back and forth and redoing prompts several times results in a lot more costs for you Cursor so why can’t I just manually select to not prune any context which is why the LLM breaks my code over and over again in the first place?
Let this cost me more, let one of these requests cost me some dollar amount, it doesn’t really matter, because progress on my product that I can show my customers and get revenue from is 100x more important than some 20 USD per month or whatever the token cost is for these requests. Price is not an issue for me, dev time is.
BTW I am doing standard React and Python stuff, the intelligence of Claude 3.6 is enough for this, that is not the issue.
We seem to be getting long context in 0.45, can’t wait to try that all though I’m certain they’ll find some way to not include crucial parts of the code anyways
Visibility on when the context has been pruned would be really useful. Having my cursor rules drop out of context and having no idea it’s happened is really frustrating.
I like how Cline shows you all the details from the requests, tokens in and out, context size (and cost)
That would be amazing but I think they would never allow that, I think they are pruning loads of what you @ in your prompt or they would lose money on every subscription. Not sure but I think somewhere above 500 rows of code it starts to prune, especially if you have multiple files, then its russian roulette but inverse if it manages to solve the problem or not. 4/5 times on issues like this it doesn’t for me and I know it would if it actually saw all the code. Thats why I’ve split my long files into 10 different smaller files and just @ the parent file + the smaller component that needs fixing
Also started to highlight code that I 100% need in the context now, or giving function and component names of relevant things instead of just @ the whole file and trust it with the pruning. Takes more time but I think its worth it in total
I understand with the lower tiers but I’m paying for Business and paying for usage on premium requests ($0.04 per request). The amount of productivity I’m getting from Cursor I would happily pay $200 a month (or more)
How disappointing. Agree, let me pay for all the token and server cost times two if it lets me get the full power… I can chose myself when I need it and don’t need it, the worst is that it decides for you
Just saw this for the first time in the Chat window. Never seen it in the composer. So guess I’m wrong about them never showing what they pruned. Can see here that basically 0% of the UID.tsx got was part of the context. UID is 600 lines, RAOP is 700 lines. I’m on 0.44.9
I can understand the need for context trimming, especially with large files. Typically, the content in very large files, like those with 1000-2000 lines, may not be strongly related, which means I need to refactor the code. Unfortunately, I’m writing Rust code, and the unit tests are in the current file, with two-thirds of the code possibly being unit tests. I can understand trimming, but it really depends on how it’s done. If it can be exposed or allow users to adjust the trimming strategy, it would provide a better experience.
On the other hand, I wrote a large number of markdown files as design documents and then organized them using NOTEPADS, but it was quite strange, as they contained a lot of design details and requirements. I can understand that if everything is given out, Claude would be able to understand my intentions. For example, I specified how to write unit tests for Redis, but with the same prompt, sometimes he can follow my intention (perhaps he trimmed the correct context), but 1/3 of the time he seems clueless, completely unaware that I had prompted him in NOTEPADS. NOTEPADS can @ files. So a large number of NOTEPADS organized together create a lot of context. I am willing to pay for this behavior because the prompts mentioned in NOTEPADS are all important, but it is evident that it has dealt me a significant blow.
