I feel like Claude 3.5 Sonnet hits the sweet spot for working with AI assistants. It works like a good assistant—helpful, focused, and not overstepping. Claude 3.7 often feels too much—it overextends or overcomplicates things at times.
If some of the smarter 3.7 features could be added to 3.5—like automatically checking for impacted files or related areas “every time” a file is modified and correcting them too—it would be perfect. That balance could make it the best for the community, my opinion.
Give this a try that @robotlovehuman posted the other day. I’ve been daily driving it for two days now, and it’s solved a good bit of my issues. All I did was drop it in the User Rules (Pervasively Rules for AI), and it works better than what I was using. 3.7 seems a bit more tame now, as well.
I really think 3.7 is very smart. However, even if I narrow it down in every single prompt with 4,5 extra hints to NOT do anything I did not want to have done, at some point your guard comes down a bit and your instructions are just a tiny bit less strict and IMMEDIATELY 3.7 jumps in as if it had waited for days for this opportunity and does 5 things hidden on the side you didn’t want it to do.
The amounts of wasted prompts with Claude 3.7 I have is still insane and it is just not worth it.
3.5 is way more on point and does not have this insane, constant eagerness to rewrite half your codebase for things you dont need.