I got the first prototype version of the code for the project I'm currently working on. It is an Arduino Sketch, that contains 1k lines, stores every single Value needed as a global Variable and every function reads and writes to these global variables, instead of taking them in as parameters. This code is so bad, that I refuse to even look at it. The effort to take the existing codebase and transform it into something useable is as high as rewriting it from scratch.
When telling my colleague that his code is absolute garbage, he just told me "its not my code, Claude did it". I am constantly pushing towards implementing the project in a proper way, but it has been nothing but fights against people who don't know how to code and who refuse to acknowledge that ignoring any standards of maintainable code does not turn out in the end and it makes sense to rewrite the project sooner than later.
I teached embedded programming for a while and students were to find out that AI often failed miserably at our tasks.
And the funny thing is that when I tried it myself to see if it’s that bad at given tasks - I explained the task like I would explain to a student, and actually got a decent result - readable comprehensive code with meaningful comments. It all simply comes down to the fact that AI won’t really help you if you can’t do without it. So I double that making the guy pass some guidelines on styling the code to Claude should at least make the code less shit to debug.
This here. It makes a huge differenche which AI you use and what you do with it.
"Free ChatGPT, make me an app" is going to result in pretty bad code.
"Claude Opus 5.5, here's a project with a well-defined agents.md, coding guidelines, tests, ..., please analyze this well-defined ticket and grill it. Together we knock out the remaining kinks in the ticket. Then put the analysis into .md files, and I review that you actually understood what you are supposed to do. Then in a new session read the .md files and start the implementation, always stopping and asking if something's unclear. Then in a new session let's review the result together, before I send it on to the next developer to do a pull request review."
This results in actually pretty good code that often catches issues that a human dev alone wouldn't.
I took an arduino class after already having programmed for years, and at one point the teacher wanted everyone to use chatgpt to write code for an assignment that was basically "run a motor while a button is pressed and stop it when it's released". The code for that is an extremely simple one liner, and if phrased correctly an llm could definitely spit out that single line, but everyone seemed to have it generate like 10+ line blocks with a bunch of if statements and stuff.