2
retoor
12h

Haha, Claude speaking the truth to me: https://devrant.molodetz.nl/preview...

Sucker.

By now, I think I can better create my own prompt system to modify source code that directly checks what exactly the changes are between the previous and current code when it comes to business logic. And a checker that actually directly checks for configuration changes.

My magic line `Do literally as i say, nothing more, nothing less` does not really work 100% with the new Claude Sonnet 4.5. I do like this version, I do not use Opus anymore, don't need it. But this one can very unexpected disappoint you.

I really question myself often, how much do they have control over how their model becomes? Is it for them a surprise as well after training? It often feels that way. Because this little flaw, that my magic sentence doesn't work anymore while being so clear, is a big failure. I am pretty sure they are aware that this model listens less good. Afaik I didn't has this issue with the previous sonnet.

Comments
  • 2
    Hmm... This video popped up for me fairly recently:

    https://youtube.com/watch/... .

    The guy had at least a few good takes, like the fact that they train those using all the code - the good && the bad. You might be seeing the results of that.

    Could be something entirely different, though.
  • 2
    yeah they don't know the result either
Add Comment