Hi,

first of all, I love spec-kit! All in all it increases the code quality a lot.

I have tried it in some large python projects - creating new features.

So, I run the normal workflow: spec -> clarify -> plan > implement.

I have a good constitution in place, lot of context and so on.

"First add tests and they must fail - because the feature isnt implemented yet"

That seems fine at the first sight. (And is a good practise) - My problem is that Claude Sonnet 4.5 implements the code as planned, but in the end, most of the tests don't match the code. Claude simply doesn't stick to it. It takes much longer to adapt the tests than to start from scratch and redo them at the end of the implementation.

Any ideas, thoughts, tips?

Best,
Georg.