Two problems in Claude's default mind model.
First problem:
The way its process works it goes like this:
1. Mock
2. Implement
3. Refine
Works great, but that gets upended on every compacting operation.
Since claude doesn't remember that the mocks are there, it will sign off on them as though you designed them.
My workaround for this is to create epics, and have the machine write detailed progress and analysis in the epic as we go.
Second problem bugs as errors:
Bugs are not errors. No seasoned developer will ever tell you that a bug is an error. You should never think of bugs as errors, because that's not what they are. It's offensive to think of bugs as errors, or failures. I will not budge on this.
Rather, bugs are proof that the structure of your app doesn't work the way you hoped it would. Nothing more, nothing less.
Every bug is a valuable data point you actually need in order to refine and ship your product.
You might not like seeing bugs. Granted.
But good QA is not for the weak of heart.
Claude does not understand this principle by default.