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.