Replying to Avatar JeffG

Been doing a very complex refactoring all week. Still an area where LLMs aren't great....

That said. Claude 4 Sonnet Thinking has been a seriously amazing pair programmer.

Always try and push the models beyond what you think they're capable of. You'll almost always be surprised.
Avatar
OceanSlim 7mo ago

I had a problem with Claude the other day where it would redo an artifact and be a few new versions but then nothing changed in the final document. I could see it making the right changes, then it spit out the old artifact with a new version on it. I haven't used thinking very much. It also seems to struggle a lot of the project knowledge base is more than 50% full or there abouts.

Reply to this note

Please Login to reply.

Discussion

Avatar
JeffG 7mo ago

Think of a model’s context window like a reverse health bar. The more it gets the dumber the model becomes.

For big repos you need to use something like Gemini with its huge context to plan work then give the smaller chunks to Claude or another model for execution.

Avatar
OceanSlim 7mo ago

Yep exactly

Avatar
franzap 7mo ago

This