When vibe coding turns into an unfixable mess

Codex
21

Before I actually want the AI ​​to help me, I need to create a backup. Make sure I thoroughly check everything the AI ​​does. Only then can I proceed to the next stage with specific checkpoints. I can do anything with the AI. Make sure I’m always creative in giving commands. Just reprimand them if they don’t do something correctly. And create a new chat tab if necessary, for example, if it has too much doctrine from previous commands. Do everything step by step:

backup→check everything→if there are no problems→backup→proceed to the next command.

But it’s not 100% AI-controlled, and it’s not yet. But at least we have what’s called an artificial assistant.

22

So Claude actually suggested this, but it seems to somewhat work:

Tell the AI to create a DESIGN.md. In it, it will record every major design decision. Don’t let the AI take complete control. Instead, tackle problems in discrete steps. Every time a step is solved, tell it to record it in DESIGN.md.

In your initial prompt, tell it that it must also read DESIGN.md every time you speak to the AI. There’s a bit of overhead in reading it, but it generally is less than the AI trying to figure out everything every time it loses context. This is also helpful when switching back and forth between different AIs.

It’s not perfect, but it is much more coherent.