Huzzah, coding from declarative pseudocode

Daniel Vaughn’s experimental editor, built on a neat reframing of what a prompt is. His complaint about current agents is that prompts are longform, imperative and transient; in Huzzah they are pseudocode, declarative and persistent. You write intent in a .hz file, saving generates the code, and editing the pseudocode later sends only the diff as the prompt so just the affected code regenerates.

The consequence is that the pseudocode file, not the chat log, becomes the durable record of intent, and it doubles as documentation that happens to be language-agnostic.

He is upfront about the limits: scaling beyond a file is untested, and it suits legacy code or unfamiliar domains poorly. The HN discussion circles the real question, which is what level of abstraction we should be working at. One good framing is that this is an ad-hoc DSL whose syntax the LLM tolerates being loose, which invites the obvious follow-up of how much ambiguity you can inject before it stops meaning anything. The sharpest objection is that this is just a spec by another name, and specs drift; the sharpest defence is that two equivalent sources, pseudocode and generated code, is exactly what makes the intent reviewable. Related, from the formal end: spec-driven development.