Andrej Karpathy asks language models to answer in the writing standard used for aircraft manuals: Reveneau

Why it mattersA one-line prompt change that routes an answer into a 900-word controlled vocabulary gives a reader less rework and a reviewer a clearer target, with no model switch and no new tooling.

One well-known researcher now gets the clearest answers out of a language model by asking it to write the way aircraft maintenance manuals are written. Andrej Karpathy, a founding member of OpenAI who joined Anthropic's pretraining team in May 2026, said on X on 1 October 2026 that he has had good results asking models to reply in ASD-STE100, a controlled form of English with about 900 approved words. Search Engine Journal's Matt G. Southern wrote the move up the following day.

Karpathy's post describes four output shapes he has had success asking for, each framed as a step up from plain prose. The first is ASD-STE100 writing. The others are diagrams, HTML pages that the reader can click around in, and custom explainer videos.

What ASD-STE100 is

The standard is maintained by ASD, which represents European aerospace, security and defence companies. Its own FAQ, cited in the Search Engine Journal piece, describes it as an international standard for technical writing, with a set of writing rules and a dictionary of about 900 approved words, most with one assigned meaning. The rules push a writer towards short sentences, one idea per sentence, and the active voice.

Karpathy wrote in his post that models already know the standard well, that its "heavy constraints on clean writing style" make the output "a lot more readable" to him, and that when the full standard reads too tight he has sometimes softened the ask to "80% of the way to ASD-STE100." A reply under the post links to an open-source agent skill that applies STE rules to text an AI agent reads, with a README noting that it is not written for marketing copy.

The ASD FAQ, as Search Engine Journal notes, says STE itself is not meant for general writing, though the underlying principles, short sentences, one topic per sentence, and active voice, carry over.

The other three shapes

Instead of writing, Karpathy wrote, asking the model for a diagram often gives a reader something "a lot easier to process, parse, and understand" when the subject has a shape. Asking for the output "in HTML" produces an interactive web page the reader can click through. Karpathy had already given similar advice in May 2026, the Search Engine Journal piece reports, about a week before he joined Anthropic, pointing at a Claude Code team practice of using HTML rather than Markdown.

The format Karpathy says he is "most bullish on" is a custom explainer video on the topic. His example prompt asks for a "3b1b style" explainer, a reference to the 3Blue1Brown mathematics YouTube channel, and instructs the model to use an ElevenLabs API key for the narration. Readers without a key can ask the model to find a free alternative that runs locally, he added.

The consequence for a team using these tools daily

Each of the four shapes is a prompt change rather than a tool change, which is the useful part for a reader on an ordinary workday. ASD-STE100 is free to download from the ASD-STAN site. A team that leaves daily reports or incident writeups to the model can test this on a current draft: paste the prose back and ask for the "80% ASD-STE100" variant, see how much rework drops out, and keep the version that reads clearer. A diagram or an HTML page is useful on a different kind of answer, which is Karpathy's second point: match the shape of the output to the shape of the question.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.