Maybe there’s a bias towards avoiding full rewrites? An “anti-refucktoring” bias
I’d be curious if the approach would be improved by having the model generate a full pelican from scratch each time and having it judge which variation is an improvement. Or if something should be altered in each loop, perhaps it should be the prompt instead
I’d be curious if the approach would be improved by having the model generate a full pelican from scratch each time and having it judge which variation is an improvement. Or if something should be altered in each loop, perhaps it should be the prompt instead