A Deeper Dive into GitHub Spec-Kit - Learning from Initial Missteps
· 12 min read
Reviewing What Went Wrong
After my first adventures with GitHub spec-kit, I was left pretty disappointed with the results. The structure and architecture looked fine on the surface, but the actual implementation was mostly placeholders. Tasks were marked as complete when they clearly weren’t, and authentication that was supposed to use Azure AD and Google was just accepting any username and password. Not good enough.
Since starting this experiment, I’ve switched to Claude Sonnet 4.5, which is a big step up from the previous model. With better prompts and a more capable LLM, I wanted to dig into why things went wrong the first time—and, more importantly, how to avoid those mistakes in future.