I've used a lot of frameworks. Spec kit is rigid, and you get out of it what you put in. It'll generate plans and tasks, but if you don't review them and iterate before executing, you're just vibe coding with extra steps. Doing it right kind of feels like accelerated waterfall.

I used GSD for a while and really liked it. It felt like a nice balance between structure and velocity.

I used it back at version 0.0.23 (around September 2025), and I just noticed I burned through a LOT of tokens, with relatively little work being done.

SOTA (state of the art) LLMs are now smart enough you can just ask it to track ADRs (Architect Decision Records - track why you did the things you did, architecture-wise).

I don't see much worth in them any more, unless you're using a local model that would still need speckit as a framework.