How to senda 4 hour long session with Claude Sonnet 3.5 missing all prompts

Yes, the previous best LLM* for coding ( I do python mostly) completely wasted my saturday for more the 4 hours with absolutely poor responses. Most of code changes created more problems that took 5 to 10 more prompts to fix or revert making me loose time and point the problem and solution multiple times.

A new annoying behavior of answering partially with a generic text ( in the lines, yes comparing text strings can cause problems like you describe, you should investigate problems in those text strings to solve ) with a worst question

**

Would you like me to explain any part in more detail?

**

asked so many times that I started including in my prompts to not ask it.

WHAT A HUGE TOKEN COST FOR NOTHING.

And finally, the code changes were so low quality that running a local model in my mac is better.

This example is the worst I’v seen :

I asked to add a few similar columns in a existing df built in a test file ( simple stuff ) to compare character unicode representations. With this code in the middle :

                rows.append({
                    'source': source,
                    'league_name': league,
                    'character': char,
                    'unicode_point': ord(char),
                    'nfc': nfc,
                    'nfkc': nfkc,
                    'nfkd': nfkd,

since the first run, I had follow up runs with this repeating errors for the same cause :

KeyError: ‘NFC’
KeyError: ‘nfc’
KeyError: ‘league_name’
KeyError: ‘league’

anyway, I would love to send, share code if that could be used to improve this
LOUSY, ANNOYING and really low quality “day” and consuming my Fast Requests for a lost saturday that I could be with my family.

  • ( now I know all of them the quality is variable, which I am quite familiar happening to human beings, and gone are the days, technology were reliable ).
1 Like

It was catastrophic Saturday for me too. Cursor or new Sonnet didn’t respond anything nothing like before. Deleting important stuff, duplicating existing functions, creating more issues than started with… It just messed up almost any file it touched. Hope for the update soon!

sorry to hear that, at least it’s not personal.

thankfully it’s not my worst nightmare “I am sorry Dave, I’m afraid I can’t do that”

1 Like

funny thing is that claude browser version seems to be way more stable than using it with cursor. The constant file deletion is so annoying

1 Like

I think this is probably one of the biggest challenges for Cursor and similar IDEs.
The longer your contexts are, the more prone they become to irrelevancy- so far, the best solution i’ve found is creating an additional file to complement your .cursorrules(but instead have this file be specific to your current session) and then start a new chat/composer.

It’s a pain to refeed relevant contexts and is sort of reminiscent of a sunken-cost fallacy: do you want to spend more time in the same session with each response deteriorating your code even more? Or do you take the time to start a new session and refresh the context with actual relevant information?

I have tried a similar approach, either with a .md mentioned with @ in the prompt or trying to use the Notepads, that I never really found meaning.

and when we start reducing our productivity to pay and use a tool for productivity, we know the answer, but I hope as an early adopter it’s the cost we pay, and we will get there.