The app I've been coding is like 95% done!
The last 5% is very difficult, my last script inflated to be too huge for the AI to handle... eventually you reach a point where any changes you make, breaks the code, and everything goes to shit. So after two complete re-writes to use the best programming practices, it's
alllllmost ready.
So, the last piece of the puzzle for my app is taking the main coordinator script and turning into a few scripts that handle the automation and use AppContext to talk to each other. That coordinator script I have that handles everything mostly works, it just gets too big and then completely shreds itself and then Claude 3.7 AI hallucinates and starts putting out gibberish, even when I switch on MAX to throw the kitchen sink at problems.
Even then, sometimes the AI just goes retarded, and you need to take a break.
Recently Cursor updated and rolled out GPT 4.1, and I've been using it and the approach seems different so far and has surprised me. I started the new chat window with a super thorough prompt that included my "cheat sheets" of what every module/plugin for my app does and the language it speaks, and a description of where I eventually see my app and how it is supposed to work. When I presented it my problem(s) in the chat window, it went on a massive dump of info with 3 possible solutions that I've never seen Claude offer before. I then had GPT 4.1 break down the options by which are the most simple yet robust and scalable, and it really smartened up.
Good thing I'm always backin' up my shit, because sometimes I get silly and just ask it to do things like speak to me in a Cajun local jive to explain the code to me and shit, and then I eventually smoke enough doobies that I forget what the software is supposed to be working on and next thing you know I'm more vibe than coding, but tonight, I am not sippin' on that whiskey, just a coke zero and a sativa brain booster.
So far tonight, GPT 4.1 has fixed several issues I couldn't fix with Claude 3.7, and testing is looking promising! The automation still doesn't work perfect but it is actually mostly working now. Before with Claude 3.7, I would get this far but none of the steps would reliably automate, now, they're mostly working. Farthest I've got yet, and 4.1 is continuing to surprise me tonight!
Just now as I was about to finish this post, GPT 4.1 fixed another issue and even suggested what the next step should be. Claude 3.7 likes to just dump code as much and as fast as it can and when it thinks it is helping you, it is sometimes adding totally random shit that you never asked for that further inflates your files. GPT 4.1 seems to work much more precisely, as seems to naturally be much more cautious. Eats up the prompts like crazy because it gives you choices and talks to you more, but what a better experience so far this evening.
4.1 is bringing me joy like biting into a Big Mac made of AI, and I'm McLovin' it!
