We hear a lot about no-code, low-code, or prompt-to-code—but these often rely on constant human input.Most agents today need us every few seconds: write a prompt, check the output, correct, repeat. You can’t step away. You definitely can’t ask them to build overnight.
Spec2Code is different. It draws from how engineers in aerospace or automotive define complex systems: with structured, layered specifications. The goal is to give agents enough information upfront that they can work autonomously for 15–30 minutes at a time—without micromanagement.
The hardest part? Writing those specs.
Nobody enjoys it, and tooling for structured requirements is basically non-existent in the software world.
So I built VibeSpec, a tool (and eventually, a framework) to turn loose ideas into structured, machine-usable specs—using agents designed specifically for that task.
Once done, the specs can be handed off to coding agents for implementation.
You can check it out here: https://vibespec.guaeca.com And here’s what the agents built from one of these specs: https:https://secretmessage.guaeca.com/
~10 minutes writing the spec with the agent ~10 minutes refining via “vibe coding” (adding missing cases) ~5 minutes testing + iterating
Would love to hear thoughts from others working on agent workflows or spec-driven development.