Jul 25, 2026·16:41·Ege Chelebi

The cure for AI slop is a 1986 aircraft manual

AIWritingClaim TestAgent Skills

16:41

Everyone is fighting AI slop one word at a time. No em dashes, stop saying delve, never open with "in today's fast-paced world." It does not work, and a thread with 2.2 million views claimed the real fix was already written: ASD-STE100, a controlled-English standard built in the 1980s so that a non-native aircraft mechanic could never misread a repair manual. Fifty-three rules, a dictionary of about nine hundred approved words, one meaning per word. Unlike "write clearly," every rule is machine-checkable.

So I mined the spec, distilled it into an agent skill, wrote a linter for the mechanical subset, and ran the test nobody had run: six realistic engineering writing tasks, four conditions, two models, scored as violations per hundred words. The banned-words list everyone uses came last. Here is the whole result:

Give a model an actual writing system and slop drops by half or more, on both models tested. The honest caveat is in the episode and in the data below: the vivid "banning words does nothing" number is Claude-specific, n is six, and STE fixes the form of writing, not whether you had anything to say.