@tante Just today I was experimenting with claude, using it to build a simple web-based game. At one point, when it implemented tests for a new set of messages, it deleted the tests for all of the rest of the message handlers. If I hadn’t been inspecting the code as it went along, I would not have noticed.
I haven’t written any code so far, but I have been inspecting the code and then providing a set of “best practices” that guide the agent to produce the code in the way that I want it. Today I added the idea that message tests should be write-once, and only modified when explicitly asked.
I suspect that people who are vibe-coding without even looking at the code can have some success, but that it leaves a lot of holes in the code that will eventually make things break in ways that they won’t be able to easily (or cheaply?) fix.