✍️STE80 Humanizer Rewrites AI Text Using 1980s Aerospace English Rules

Your AI-written docs are about to read like a person wrote them

Somebody finally fixed the worst part of LLM output: it reads like a press release from a robot. STE80 Humanizer strips the AI tells and rebuilds sentences using ASD-STE100, the controlled English used in aircraft maintenance manuals. The goal is text people can read without effort, and dodging AI detectors doesn't enter into it. If you're reviewing agent output, PRs descriptions, or generated docs, this is for you. Karpathy's argument from October 2, 2026 is that as LLMs do more of the work, people spend more time reading what the models produce. Ask for an ASD-STE100 explanation and you get short sentences, one instruction per sentence, one meaning per word. Less to parse, less to misread. The skill runs in three modes: Humanize (default), Light, and Strict STE. Strict applies the 100% rule, which bans words like "may" and "should" and limits "check" and "test" to nouns. Default 80% keeps your full structure and only trims vocabulary that stops sounding natural. It keeps "may" as "may", keeps "40+" as "40+", and flags contradictions in the source instead of quietly fixing them. Edits are numbered (§1 to §26 for AI tells, S1 to S15 for structure), so you can audit every change.