Had almost 2 weeks loving Claude in Composer mode.
Made massive progress on an app (as a non developer).
Then as key files started getting too big, (2000 lines) I asked it to turn them into components, from that point I’ve been endlessly going in circles getting Claude to fix Claudes crappy code at every new prompt.
I’m aware there’s a context window. I regularly start new chats, and have created both an architecture doc and a UI/human centered set of guidelines. But Claude is currently operating like a blind goldfish with alzheimers. Endlessly creating quick fixes, duplicating features and not solving root problems its creating. Creating new issues with every prompt and going in circles.
It seems totally unable to hold the context of the files or structure it created.
I started trying to force it to read all relevant files before making any changes, and while trying to get to the bottom of this it came back with this reply that I thought was interesting and maybe part of the problem - it only ever seems to be reading 250 lines of any relevant file.
I tried to prompt around this, also added it to cursors ai instructions, but it won’t remember it, seems hard coded. Cursor said…
"I’m only reading lines 1-250 of files because of a limitation in the read_file tool. According to its description:
“Note that this call can view at most 250 lines at a time.” "
Is that a Claude or Cursor rule? Either way it seems like a crazy limitation and its own type of context window. How can you write or fix code, checking multiple files if you only ever read the first 250 lines of anything?
After my initial excitement it feels like I’ve hit the limits of its capabilities at this point, at least in the hands of a non developer?
That’s a valid point, consequently, Cursor’s composer may not function optimally in larger projects, especially when handling substantial tasks across multiple files.
at this point i’m starting a new chat each session, and forcing claude to read key files, also have to force it to read the whole file a lot. and when I see it going into a loop I switch to a new chat with gpt. gpt hates looking at files even more, its like a cocky lazy reluctant developer, but once forced is often better at solving root cause problems. i then take the solution back to claude who is more pleasant to work with. lol
I spent $303 and within a month my account stopped working. I have contacted [email protected] but it has been many days with no resolution.
please take a look for my account [email protected] best。
I’ve noticed Claude has a real tendency towards quick short term fixes, often suggesting ridiculous solutions that even as a non dev I can spot. Regularly forcing it to “fix the root cause” delivers completely different results. chatgpt seems better at finding root causes, but feel like an arrogant ■■■■■ to work with.