Claude 3.7 Review (Is This Real for Coding?)
Starting today, I ran Claude 3.7 in coding mode for real work, and it hit me that we've entered an era where you can have AI analyze legacy code.
What really blew my mind was how its refactoring suggestions are on a whole different level compared to GPT-4.0. It automatically follows the commit convention, writes test code on its own... The only problem is it burns through tokens fast lol.
Of course, it's not perfect—there were cases where it wrote code based on wrong assumptions, so human review is a must. Still, for personal projects, I now feel like developing without AI is actually the inefficient choice.
6 answers
Agreed, Claude coding is really different.
So what if token consumption is fast... productivity goes up so much.
Well, I'm not so sure about that. I had an agent implement OAuth authentication, but it kept going in the wrong direction and just kept asking questions, so I ended up fixing it all myself. A final review is definitely necessary, but for short projects, that 'final review' just ends up being a review of everything.
Legacy analysis feels like reading through a contract, lol. I also got chills yesterday when I saw it clean up 3,000 lines of unused code right away.
That still has limitations in large projects, doesn't it? I'd agree when it comes to personal or toy projects, but in a real-world codebase with tangled multi-module dependencies, GPT and Claude both fumble around in similar ways.
Got a source? Even when I used Claude 3.5, it just made pretty refactoring suggestions but the behavior stayed the same... Is 3.7 any different? I'm wondering if it's worth trying out for code review these days.