Founders building with AI: at which feature did your app start breaking itself?
The demo works on Saturday. By feature four, every new prompt breaks something that used to work.
I've heard that story from founder after founder, and it's never the model's fault. Nothing ever wrote down what the product actually is. No requirements, no spec, no definition of done. So every prompt is a fresh guess, and guesses don't add up to a system.
Senior engineers have always started from the spec. Most people building with AI skip it, because no tool made them do it.
So, for everyone here who has shipped something with AI:
At which feature did your build start breaking itself?
What fixed it: a written spec, tests, starting over, something else?
If you write requirements first, what goes in them?
Disclosure: I'm the founder of Codalio and I've spent seven years on this problem. We launch Codalio here tomorrow. I'm not asking for votes, only your stories, and I'll reply to every one.
Replies