Here is a sentence an AI will happily write for your company: “Our tailored solution helps teams get more done with the latest technology.” (An illustration I made up. It is not from a client.) Nothing in it can be checked. Nobody could tell you whether it is true.

Now the version with a rule behind it: “The system sends the quote to the customer when the salesperson approves it.” Short. One action. You can test it.

The gap between those two sentences is what Andrej Karpathy was poking at on October 2. In a post of tips on understanding what language models write, he said to ask your LLM to explain something in ASD-STE100, “a controlled language specification originally developed for aerospace maintenance documentation.” Sometimes he softens it to “80% of the way to ASD-STE100,” because the spec is stringent.

One caveat first. Karpathy was talking about reading model output more easily. He was not talking about marketing. What follows is my extension, and I flag it as mine.

What the standard is, per the people who maintain it

Aakash Gupta added a backstory in a thread. I went to the source. The maintenance group for ASD-STE100 says Simplified Technical English was developed in the late 1970s, at the request of European airlines, so that maintenance documentation would be easier for readers with only basic English. The first guide came out in 1986. Issue 9 is dated January 2025.

Wikipedia’s summary puts it at 53 writing rules and a dictionary of about 900 approved words, each with one meaning and one part of speech. Instructions run at most 20 words per sentence, descriptions at most 25. One instruction per sentence, one topic per paragraph, six sentences per paragraph at most. An official copy is free.

Gupta’s thread also cites a list of about 1,200 banned words and a few specific examples. I could not tie those to a source, so I am leaving them out.

Vague AI copy is a spec problem

People usually ask a model for text like this: “professional but approachable tone, results-oriented.” No rule can check that. Two people on the same team would disagree on whether a draft meets it.

What makes STE useful is that most of its rules can be checked mechanically. Does the sentence run past 20 words? Yes or no. Did you use “customer” to mean two things in one text? Yes or no. A model can follow a rule you can count, and a script can verify it, instead of your judgment at eleven at night.

One measurement points the same way. AI engineer Ege Chelebi posted that he turned ASD-STE100 into an agent skill plus an automated checker and tested it against the popular fixes on two models. By his test, giving the model a real writing system cuts slop by half or more, while the banned-words list everyone uses cut it only 3% on Claude. He says himself that it fixes form, not substance. It is one person’s test with his own definition of slop, and I have not replicated it. I take it as a hint.

It is the same logic as the AI writing policy I wrote about on September 30, one level down. The policy says who signs and what AI may draft. The style spec says what a draft has to look like before it deserves to reach the person who signs.

What to borrow and what to leave

STE was written so a mechanic can follow a procedure without misreading it. It fits wherever the reader has to do something right or understand a description without ambiguity: onboarding emails, product sheets, FAQs, the scope section of a proposal, support macros.

It does not fit a headline or your brand story. There you want voice, which is why Karpathy stops at 80%. It is also English only. A team writing in Spanish should take the principles and write its own version.

Polishing with AI afterward flattens voice, according to Nature. The spec is the opposite move: constrain the draft up front.

A one-page spec you can write this afternoon

Collect twenty pieces of yours that worked and measure their sentence length. That sets your limits from what you already do well, not from someone else’s number. Then write rules like these. They are my example, so adjust them:

1. One meaning per term. “Lead” is a person who left their details.
“Opportunity” is a lead with a meeting booked. Never swap them.
2. One name per thing. If paragraph one says “platform”, do not
write “solution” or “tool” for variety.
3. 20 words maximum per sentence in instructions, 25 in descriptions.
4. One instruction per sentence. One topic per paragraph.
5. Active voice. The subject is the person or the company, not “the system”.
6. Every claim carries a number, a name or a date. Otherwise cut it.
7. Banned: “comprehensive”, “cutting-edge”, “leverage”, “solutions”, and the
ten filler words your industry overuses. This is the rule that weighs least.

Rule 1 is the hard one, because it forces you to decide what each term means in your business. At IQ Source the definitions audit in AI Maestro does the same job: it finds where one word means two things in two departments. If two people on your team define “opportunity” differently, your AI will write with both definitions.

Then test it on ten pieces. Count how many rules each draft breaks, using a plain script for sentence length and the banned list. It measures itself, with no person reading.

A spec does not replace human review, and a text can follow all seven rules and still be dull. I treat it as a floor. Everything above the floor belongs to whoever signs.

If your team already generates content with AI and the complaint is always the same, “it sounds generic,” let us write the first version of your spec with you