Karpathy's aerospace-English trick, packaged as an agent skill
A two-pass humanizer that strips the 26 known AI tells, then rewrites your text 80% of the way to ASD-STE100 — the controlled English of aircraft maintenance manuals.

What it does
STE80 Humanizer is an agent skill (Claude Code, Codex, Cursor, and others via the skills CLI) that rewrites AI-generated text so it reads like a careful person wrote it. It runs two passes: first it removes the 26 tells catalogued in Wikipedia’s “Signs of AI writing” and blader/humanizer — the “not just X, it’s Y”, the “Great question!”, the stacked hedges. Then it rebuilds the sentences 80% of the way to ASD-STE100, the controlled English invented in the 1980s for aerospace mechanics whose first language wasn’t English, where a misread step can kill someone.
The interesting bit
The insight is that most humanizers only swap vocabulary. The text loses “seamless” but keeps 40-word sentences carrying three ideas in passive voice — it still smells like AI. Pass 2 attacks structure instead: one idea per sentence, active voice, condition first, 20–25 word limits. The “80%” matters because full STE is stiff — at 100%, “check” and “test” are nouns only and “may” is banned. At 80% you keep the sentence discipline and lose the vocabulary straitjacket.
Key highlights
- Two numbered rule sets in
SKILL.md: §1–§26 for AI tells, S1–S15 for STE80 structure, each with before/after examples - Three modes: default humanize, light touch (tells only), and strict STE for procedures and safety notes
- Fidelity-first: keeps every fact, preserves “may” as “may”, flags contradictions in the source instead of smoothing them over
- Can match your own voice if you supply a writing sample
- Reports edits by rule number ("§1, §8, S2") so you can audit what changed
- The README even fact-checks the STE overview graphic from Karpathy’s post against Issue 9 of the spec
Caveats
- Explicitly not aimed at beating AI detectors — the goal is readability, and the README says so twice
- The “clearest difference from a plain prompt” claim rests on the authors’ own tests; no external benchmark is given
- Karpathy’s post is dated 2 October 2026, which is either a typo or a time traveller — the README doesn’t flag it
Verdict
Worth installing if you regularly clean up model output for docs, runbooks, or replies to humans. If you were hoping for a detector-evasion tool, this is politely not that.
Frequently asked
- What is Dodaios/ste80-humanizer?
- A two-pass humanizer that strips the 26 known AI tells, then rewrites your text 80% of the way to ASD-STE100 — the controlled English of aircraft maintenance manuals.
- Is ste80-humanizer open source?
- Yes — Dodaios/ste80-humanizer is open source, released under the MIT license.
- What language is ste80-humanizer written in?
- Dodaios/ste80-humanizer is primarily written in Python.
- How popular is ste80-humanizer?
- Dodaios/ste80-humanizer has 1.1k stars on GitHub.
- Where can I find ste80-humanizer?
- Dodaios/ste80-humanizer is on GitHub at https://github.com/Dodaios/ste80-humanizer.