Write It in ASD-STE100

Andrej Karpathy posted a short guide on X recently about how to read language-model output as the models take on more of the work. His first tip is a writing constraint: ask the model to explain the topic in ASD-STE100.

ASD-STE100 is a controlled language from the aerospace and defence industry. It was built for maintenance documents, where a wrong reading can cause a wrong action. Writers use an approved word list. Each word has one meaning. Sentences stay short, usually under 20 words. The text uses the active voice and simple verb forms. Synonyms are not allowed for the same idea.

Karpathy says models already know this specification, and the constraints produce cleaner prose than a normal explanation. The full spec is strict. If the result feels too flat, he suggests asking for “80% of the way to ASD-STE100.”

That request is useful when your job is to check the answer, not just to receive it. Short sentences make the claim easy to see. They also make the limit of the claim easy to see. Stable wording makes two explanations easier to compare.

He does not stop at writing. In the same post he ranks diagrams above prose, HTML pages above diagrams, and custom explainer videos above pages. ASD-STE100 is the first step on that ladder: limit the language, then move the same idea into a form you can scan faster.