Though one thing I have to say: I’m very annoyed by it’s constant agreeing with what I say, and enabling me when I’m doing dumb shit. I wish it would challenge me more and tell me when I’m an idiot.
There’s a balance to be had there, too… I have been comparing a few AI engines to compare their code generation capabilities. If you want an exercise in frustration, try to make an old school keypress driven application on a modern line-oriented terminal interface while still using the terminal for standard text output. I got pretty far with Claude, then my daily time limits were kicking in. Claude did all that “you’re so right” ego stroking garbage, but also got me near to a satisfactory solution. Then I moved into Google AI and it started out with reading my the “you just can’t do that, it won’t work” doom and gloom it got from some downer stack overflow or similar material. Finally, I showed Google my code that was already doing what it was calling impossible and it started helping me to polish the remaining rough spots. But, if you believed its first line answers you’d walk away thinking that something relatively simple was simply impossible.
Lately, I have taken to writing my instructions in a requirements document instead of relying so much on interactive mode. It’s not a perfect approach, but it seems to be much more stable for “larger” projects where you hit the chat length limits and have to start over with the existing code - what you’ve captured in requirements tends to stick around better than just using the existing code as a starting point of how things should be then adding/modifying from there. Ideally, I’d like it if the engine could just take my requirements document and make the app from that, but Claude still seems to struggle when total LOC gets into the 2000-5000 range for a 200-ish lines requirement spec.
There’s a balance to be had there, too… I have been comparing a few AI engines to compare their code generation capabilities. If you want an exercise in frustration, try to make an old school keypress driven application on a modern line-oriented terminal interface while still using the terminal for standard text output. I got pretty far with Claude, then my daily time limits were kicking in. Claude did all that “you’re so right” ego stroking garbage, but also got me near to a satisfactory solution. Then I moved into Google AI and it started out with reading my the “you just can’t do that, it won’t work” doom and gloom it got from some downer stack overflow or similar material. Finally, I showed Google my code that was already doing what it was calling impossible and it started helping me to polish the remaining rough spots. But, if you believed its first line answers you’d walk away thinking that something relatively simple was simply impossible.
Lately, I have taken to writing my instructions in a requirements document instead of relying so much on interactive mode. It’s not a perfect approach, but it seems to be much more stable for “larger” projects where you hit the chat length limits and have to start over with the existing code - what you’ve captured in requirements tends to stick around better than just using the existing code as a starting point of how things should be then adding/modifying from there. Ideally, I’d like it if the engine could just take my requirements document and make the app from that, but Claude still seems to struggle when total LOC gets into the 2000-5000 range for a 200-ish lines requirement spec.