On October 2, 2026, Andrej Karpathy — OpenAI founding member, former Tesla AI lead, now on Anthropic's pretraining team — posted a tip on X that sounds absurd on its face: ask your language model to explain things in ASD-STE100, a controlled form of English originally built to make aircraft maintenance manuals readable. His claim is that the output is "a lot more readable" than what you normally get. He's right, and the reason it works tells you something important about how to actually use these models.

ASD-STE100 stands for Simplified Technical English. It's an international standard maintained by ASD, the European aerospace and defense trade association. The standard enforces a set of writing rules — short sentences, one topic per sentence, active voice, approved vocabulary of roughly 900 words where each word has one meaning — that were designed in the 1980s to stop mechanics from misreading maintenance instructions and crashing planes. It's the most boring-sounding writing standard in existence, and it's one of the most effective prompt engineering techniques anyone has published.

Why controlled English works

The reason STE produces better LLM output isn't mystical. It's mechanical. When you ask a model to write in ASD-STE100, you're imposing constraints that eliminate the exact failure modes that make LLM text frustrating to read.

Standard LLM output tends toward three problems: ambiguity, verbosity, and false fluency. The model uses words with multiple meanings and expects you to figure out which one from context. It pads sentences with qualifiers and subordinate clauses that dilute the actual point. And it writes with a confident, polished tone that makes bad information sound authoritative. STE attacks all three.

The vocabulary restriction kills ambiguity. When the model can only use words from an approved list where each entry has a single definition, it can't hide behind vague language. "Utilize" becomes "use." "Facilitate" becomes "help" or a more specific verb. "Implement" becomes whatever the model actually means — "install," "set up," "run," depending on context. The constraint forces the model to pick the right word instead of the impressive-sounding one.

The sentence structure rules kill verbosity. STE limits sentences to roughly 20 words for procedures and 25 for descriptions. Each sentence gets one topic. No stacking three dependent clauses into a single 60-word monster that buries the verb. The model has to break complex ideas into sequential statements, which is how humans actually process technical information.

The combination kills false fluency. When the model is constrained to simple words and short sentences, it's much harder for it to produce the kind of confidently wrong paragraph that sounds like a senior engineer wrote it but says nothing. The constraints make vagueness structurally difficult.

Karpathy's four formats

Karpathy didn't stop at controlled English. His post walked through four output formats he's been experimenting with, each one building on the last.

The first is STE — text, but disciplined text. The second is diagrams, which he calls "a lot easier to process, parse, and understand" for many topics. The third is HTML — asking the model to output an interactive web page instead of a markdown document. He credited Anthropic's Claude Code team with popularizing this approach, and the Claude Blog published a detailed piece on it in May 2026. The fourth, which he says he's "most bullish on," is custom explainer videos. His example prompt requests a "3Blue1Brown style" animated explanation and uses ElevenLabs for narration.

The throughline across all four is the same idea: the raw text output of a language model is the default, not the ideal. Most people ask a question and accept whatever format the model gives back. Karpathy's argument is that you should be deliberate about the format, because the format determines how much of the model's knowledge actually reaches your brain.

What this means in practice

You don't need to memorize the STE standard to benefit from this. Karpathy acknowledged that STE's constraints are "quite stringent" and said he sometimes asks for "80% of the way to ASD-STE100" instead. An open-source skill applying STE rules to AI agent output was already on GitHub within hours of his post.

The practical version is simpler: tell your model to use short sentences, active voice, and simple words. Tell it to avoid jargon unless it defines the jargon first. Tell it to break complex explanations into numbered steps. These are the principles behind STE, and they produce noticeably better output whether you're asking about code, business strategy, or how to fix a sink.

The HTML suggestion is worth trying separately. Instead of asking for a markdown explanation, ask for a self-contained HTML page with inline CSS. You get a visual layout, interactive elements if relevant, and a format that's easier to scan than a wall of text. For anything involving comparisons, timelines, or structured data, HTML beats markdown every time.

The bigger shift

Karpathy closed his post with a line worth sitting on: "a lot more of our work will rise up the abstractions into oversight and understanding." He's saying that as models get better at generating content, the human job shifts from producing text to reviewing and directing it. The person who knows how to ask for the right format, in the right constraints, gets better results than the person who just types a question and hopes.

This is already happening. The developers getting the most out of tools like Claude Code, Cursor, and GPT-6 aren't the ones with the best raw coding skills. They're the ones who've figured out how to structure their requests — what context to provide, what format to demand, what constraints to impose. Prompt engineering isn't a meme. It's the difference between getting a polished wall of text and getting something you can actually use.

Aircraft maintenance manuals are readable because someone decided that clarity was more important than sounding smart. Your LLM outputs could use the same philosophy.

Sources: Andrej Karpathy on X, Search Engine Journal — OpenAI Founding Member: Make LLMs Write Like An Aircraft Manual, ASD-STE100 FAQ