Andrej Karpathy, one of the founding members of OpenAI, has suggested asking language models to write their explanations in ASD-STE100, the Simplified Technical English used to write aircraft maintenance manuals. He made the point on X in early October as part of a broader argument: as AI does more work on its own, people will spend most of their time reviewing and understanding what it produces, and the long, padded prose models tend to generate is a poor format for that job.

Key points on Karpathy’s ASD-STE100 suggestion in 30 seconds

The idea: ask AI models to write in ASD-STE100, the controlled English used in aerospace maintenance documentation.
The standard: maintained by the European association ASD, it has about 900 approved words and 53 writing rules; the current release is Issue 9, from January 2025.
Beyond text: Karpathy ranks diagrams, interactive HTML pages and generated explainer videos above it.
Why it matters: if AI produces more of the work, reviewing it quickly and without ambiguity becomes a core task for technical teams.

According to Search Engine Journal, Karpathy said he had good results asking models to explain things in ASD-STE100 because the standard places heavy constraints on writing style. He admitted it is quite stringent and that, when he relaxes the request, he settles for text that gets roughly 80% of the way there. For sysadmins the idea has an obvious application: much of what people now ask an AI assistant for is procedures, error explanations, incident summaries and internal documentation, exactly the kind of text where ambiguity costs time or worse.

An aviation standard with 900 words and 53 rules

ASD-STE100, known as Simplified Technical English (STE), came out of the European aerospace industry so that maintenance manuals could be understood with no room for interpretation, including by technicians whose first language is not English. The group that maintains it, the STEMG, was set up in 1983 within the association now known as ASD (Aerospace, Security and Defence Industries Association of Europe), and the first official guide, then called the AECMA Simplified English Guide, was published in 1985. Some accounts place its origins in the late 1970s; the reference documentation dates those two milestones to the 1980s.

The current version is Issue 9, released in January 2025, and the standard has been a free download from the ASD website since Issue 6 in 2013. It combines two parts: a dictionary of about 900 approved words, most with a single meaning, and 53 writing rules. Among the best known are a 20-word limit for procedural sentences and 25 words for descriptive ones, one instruction per sentence, active voice, and no switching terms for variety: if a part is a bolt, it stays a bolt throughout the document. The standard is no longer just about aircraft. According to the Issue 8 distribution log, 64% of those who downloaded it worked outside aerospace and defence, in fields such as automotive, energy and medical devices.

Two caveats are worth keeping in mind. ASD itself says in its FAQ that the standard is not intended for general writing, although principles such as short sentences, one topic per sentence and active voice may carry over, and that no tool can replace knowing it. After Karpathy’s post, several agent skills that apply STE rules circulated on GitHub, but at least one of the most widely shared does not include the dictionary, which is half of the standard.

From text to throwaway HTML: the format ladder

Karpathy’s suggestion goes beyond prose. In his post he ranked several formats from least to most useful for understanding what a model produces: first, controlled text in ASD-STE100; then diagrams, which he finds “a lot easier to process, parse, and understand”; then interactive HTML pages; and at the top, explainer videos in the style of 3Blue1Brown generated by the model itself, with narration from a service such as ElevenLabs or a free local alternative. That last format is the one he said he is “most bullish on”. He closed by encouraging people to ask for “large, custom, discardable software artifacts”.

The underlying idea is that generating code is now cheap enough to justify building a tool just to understand something and deleting it afterwards. In a post on Medium, developer Nikhil describes trying both recommendations for a week. To understand how the learning rate affects gradient descent, he asked for a single HTML file with a slider and a ball rolling across an error surface; by his account, a few seconds of moving the slider taught him what would previously have taken several paragraphs, and then he deleted the file. He also writes that adding a directive to apply the spirit of ASD-STE100 to his technical prompts made the answers read more like a workshop manual than a chat.

What an operations team can take from it

Runbooks, change procedures and post-incident reports share a problem with aviation manuals: someone has to follow them under pressure without second-guessing what each step means. Asking the AI to write them with one action per line, sentences under 20 words and the same name for each server, service or volume from start to finish cuts down on double readings. The same applies to reviewing what an agent proposes to do: a list of short steps with fixed terms is easier to audit than three paragraphs in which the same disk appears as a volume, a drive and storage.

The higher rungs fit daily work too. A Mermaid architecture diagram or a generated map of a codebase, like the one Graphify builds on top of Claude Code, explains a topology better than a description, and a disposable HTML page with controls can show how latency or memory use changes as a parameter moves. It also fits practices many teams already use, such as setting writing rules for the agent in CLAUDE.md. Karpathy’s suggestion is a usage tip, not a feature of any product, and how well it works depends on the model and the task, but it has the advantage of being testable in the very next prompt.

Frequently asked questions

What is ASD-STE100?

It is Simplified Technical English, a controlled form of English created by the European aerospace industry to write technical documentation without ambiguity. It has about 900 approved words and 53 writing rules, and the current version is Issue 9, from January 2025.

What did Andrej Karpathy suggest?

Asking AI models to explain things following ASD-STE100 and, when text is not enough, asking for diagrams, interactive HTML pages or custom-generated explainer videos.

What are the main ASD-STE100 writing rules?

Among the best known: procedural sentences of no more than 20 words, descriptive sentences of no more than 25, one instruction per sentence, active voice and consistent terminology.