Over the past four years, I've shipped a lot of vibe-coded software. It feels great in the moment — but keep vibe coding the same project on and off for three to six months, and the technical debt piles up. Things start falling apart at the seams, and you end up playing whack-a-mole with bugs popping up in unexpected places.
So between Andrej Karpathy's vibe coding and Professor Andrew Ng's more rigorous AI Engineering, there has to be a middle ground. I call it spec vibing.
Enter speccheck — a set of agentic skills that:
* write, review, and revise the spec (the document you hand your coding agent)
* plan the build in waves (small, verifiable increments rather than one big build)
* build against the spec
* then run speccheck as a final gate before you call it done
speccheck answers one question: do you have evidence that what the coding agent built fully satisfies the spec? It can check this two ways — mechanically (grep-like, fast but not very reliable), or using LLM-as-judge (local or remote). The coding agent iterates until your code is speccheck clean.
Over the past four years, I've shipped a lot of vibe-coded software. It feels great in the moment — but keep vibe coding the same project on and off for three to six months, and the technical debt piles up. Things start falling apart at the seams, and you end up playing whack-a-mole with bugs popping up in unexpected places. So between Andrej Karpathy's vibe coding and Professor Andrew Ng's more rigorous AI Engineering, there has to be a middle ground. I call it spec vibing. Enter speccheck — a set of agentic skills that:
* write, review, and revise the spec (the document you hand your coding agent) * plan the build in waves (small, verifiable increments rather than one big build) * build against the spec * then run speccheck as a final gate before you call it done
speccheck answers one question: do you have evidence that what the coding agent built fully satisfies the spec? It can check this two ways — mechanically (grep-like, fast but not very reliable), or using LLM-as-judge (local or remote). The coding agent iterates until your code is speccheck clean.
The introduction: https://rioffe.github.io/speccheck