Ask HN: How would you lint natural-language specs before LLMs run them?
We lint code before execution—what’s the practical equivalent for natural-language specs before an LLM/agent runs them? I’m not looking for “NL ≠ code” debates; I’m after concrete designs, rules, and tooling that catch mistakes pre-execution.
Comments
I would say the most important thing for linting natural language specifications would be a strictly enforced, well defined grammar that precisely defines what words could be used in what order, do that the linter could identify tokens that were out of place. Otherwise, you have ambiguity that of all too common in natural language.
But you also said you don’t want to discuss why natural language isn’t code.
Rex