Andrej Karpathy said in a post on X Thursday that he’s had success asking language models to explain things in ASD-STE100, a controlled form of English built to make aircraft maintenance manuals easier to read.
Karpathy, a founding member of OpenAI, joined Anthropic’s pretraining team in May. The post moves through four formats for reading model output, from controlled writing to diagrams, HTML pages, and explainer videos, calling each step “even better.” He wrote that as models do more of the legwork, “a lot more of our work will rise up the abstractions into oversight and understanding.”
Karpathy’s Case For Controlled English #
Karpathy wrote that models know ASD-STE100 well and that its “heavy constraints on clean writing style” produce text he often finds “a lot more readable.” Because the standard is “quite stringent,” he has sometimes softened the request to “80% of the way to ASD-STE100.”
ASD, which represents European aerospace, security, and defense companies, maintains the standard. Its FAQ calls Simplified Technical English (STE) an international standard for technical writing, with a set of writing rules and a dictionary of about 900 approved words, most with one meaning.
A reply under the post links to an open-source skill that was already on GitHub and applies STE rules to text that AI agents read. Its README says the skill isn’t built for marketing copy.
Diagrams, HTML Pages, And Explainer Videos #
Instead of writing, Karpathy suggests asking for a diagram, which he says can be “a lot easier to process, parse, and understand.”
He also recommends asking for output “in HTML” to get an interactive web page. He gave similar advice in May, about a week before he joined Anthropic, when he shared an X post by Anthropic’s Thariq Shihipar on why members of the Claude Code team use HTML instead of Markdown. A version of the piece ran on the Claude Blog later that month.
The format he is “most bullish on” is custom explainer videos on any topic. His example prompt asks for a “3b1b style” explainer, a reference to the 3Blue1Brown math videos, and tells the model to use an ElevenLabs API key for the narration. He added that people without a key can ask the model to find free alternatives that run on their own computer. “This is actually starting to work!” he wrote.
Why This Matters #
Karpathy’s four suggestions are different ways to change how model output is presented to the person reviewing it. He frames the writing, diagram, and HTML formats as requests to a model, while his video example also needs an ElevenLabs API key or a free local alternative for narration. For the controlled-English tip, the standard’s own FAQ says STE isn’t meant for general writing, though principles like short sentences, one topic per sentence, and active voice can carry over.
Looking Ahead #
Karpathy closed the post by urging people to ask models for “large, custom, discardable software artifacts” that “would have never made sense to create before.”
The group behind ASD-STE100 gives out the full standard for free, and its FAQ says no tool can replace it. The open-source skill linked in the post’s replies says it doesn’t include the standard’s dictionary.
Featured Image: Teodora ART/Shutterstock