Agreed, also I'll summarize my experience with Spec-kit + Claude so far (project is default node+react fullstack):

1. Does not actually follow constitution, or does that occasionally here or there. Example is that i have an article stating that all features should be actually tested (really RUN the tests) but it does not do that.
2. In plan & tasks phase it creates a whole lot of unused (or unusable) tasks, contracts, endpoints (whatever I'm working on)
3. /implement does not actually finish all of the tasks, though marks them as "complete", and i have to push it with "make sure you finished..."
4. It ignores project limitations and previous implementation specs, although my "constitution" clearly states to act otherwise
5. It astonishingly over-engineers almost all of solutions it comes up with during research, ignores ready-made library uses, skips already implemented things.
6. I tried both small-sized (like 1 hour estimate) and medium-sized (8-hour) tasks | features, and problems above are actual disregarding task size.

As a result, after "completing" a task with spec-kit I have to re-evaluate every task, and basically go with "make sure it actually works and is tested properly..."

Maybe I'm using it wrong, but at this point I'm removing spec-kit from my AI-slop-projetcs, and switch to regular vibing with Claude.

I took a note from the Let It Cook episode. Den said at the time that sometimes you need to simply reply, "Check off the completed tasks and continue". I find myself doing this bit in Copilot CLI also to have the coding complete 100%