Andrej Karpathy has a new prompting trick for clearer AI output: ask the model to write like it is producing aircraft maintenance documentation.
In a post on X on Thursday, the founding OpenAI member and current Anthropic pretraining team member said he has had success asking language models to explain things in ASD-STE100, a controlled form of English created to make aircraft maintenance manuals easier to read. His argument is simple: the standard’s heavy constraints on clean style produce text he often finds “a lot more readable.”
What ASD-STE100 actually is
ASD-STE100 is Simplified Technical English, an international standard for technical writing maintained by ASD, the group representing European aerospace, security and defense companies. It combines strict writing rules with a dictionary of about 900 approved words, most of them carrying a single meaning.
Karpathy says models already know the standard well. Because it is stringent, he sometimes softens the request to “80% of the way to ASD-STE100” to keep the output useful without overcorrecting.
Why output format is a marketing lever
Most prompt engineering focuses on what you ask. Karpathy is pointing at something else: how the model presents the answer. A dashboard-style HTML page, a clean diagram or a tightly written summary can turn an unusable wall of text into something a team can review in seconds.
For marketers, this is not a license to write customer-facing ads in aircraft English. The standard’s own FAQ says it is not meant for general writing. But the principles are portable: short sentences, one topic per sentence, active voice. They are ideal for internal briefs, AI agent instructions and reporting.
- Controlled English: ask for ASD-STE100 or an 80% version for internal docs and agent prompts.
- Diagrams: request a visual explanation when a process, funnel or workflow feels heavy.
- HTML: ask for an interactive page you can click, filter and explore instead of read.
- Explainer video: ask for a “3b1b style” video, with narration through an ElevenLabs API key or a free local alternative.
From text to oversight
Karpathy’s post moves through four formats, calling each step “even better.” The point is not to consume more media. It is to review more efficiently. If a model can generate a diagram or an interactive page, the human can spend less time reading and more time judging the idea.
He frames it as a shift in work: as models do more of the legwork, “a lot more of our work will rise up the abstractions into oversight and understanding.” That is a useful way to think about AI in marketing teams: writing is becoming cheap, but clear review and good judgment are not.
The bigger shift
Karpathy’s most bullish format is custom explainer videos, using a 3Blue1Brown-style prompt and ElevenLabs narration. He said this is “actually starting to work.” His closing advice is to ask models for “large, custom, discardable software artifacts” that would never have made sense to create before.
The takeaway for growth teams: treat output format as a prompt variable. A controlled-language layer can quickly improve the thing models actually hand you, without waiting for the next model release.
Source: Search Engine Journal
