AI NewsProductivityAnnouncement
STE80 Humanizer rewrites AI text against an aerospace plain-English standard
STE80 Humanizer is an MIT agent skill that rewrites AI-written text against ASD-STE100, an aerospace plain-English standard. The Show HN post has 67 points and the repository ranks 18 on Trendshift's daily list.
Image: GitHub
Why it mattersThe standard was written for safety-critical manuals, so it bans the vague hedging that marks AI prose. A rewrite that targets it keeps the meaning and cuts the tells.
An aircraft repair manual cannot afford a sentence a mechanic reads two ways. The aerospace industry wrote a standard for that problem in the 1980s called ASD-STE100, Simplified Technical English. It picks around 900 approved words. It limits a sentence to 20 words for instructions. It bans a vague verb like "could". A new agent skill borrows the standard to clean up text an AI wrote.
Dodaios/ste80-humanizer is the project. It is released under the MIT licence. It had 141 stars on GitHub by the time the Show Hacker News post reached 67 points. Trendshift ranked it 18th on its daily list.
What the skill does in two passes
The first pass removes 26 "tells" that the author lists from a Wikipedia catalogue of signs of AI writing. A hedge like "could potentially" gets cut. A promotional word like "leverage" or "unlock" gets replaced. The second pass restructures the sentences toward ASD-STE100. The author targets 80 percent compliance, which is where the name comes from. The remaining 20 percent keeps the text natural: common words and polite modals like "may" and "should" stay in.
The README gives a short example. The AI line "could potentially cause cascading restarts" becomes "can break". The point is that the before and the after mean the same thing. The first version was only pretending to be cautious.
Why the standard works on AI text
ASD-STE100 was written to stop a reader from misunderstanding a step that could cost a life. Its rules target exactly the writing habits AI models have most trouble unlearning: passive voice, multi-clause sentences, nominalised verbs, abstract nouns. It picks one word for each concept, so "test", "check" and "verify" collapse to one approved verb. A rewrite that targets the standard pulls the sentence back to something a non-native reader can follow.
How to use it
The project is a Claude Code skill. The command is /ste80-humanizer followed by the text. A plain-English instruction also works. One example in the README is "Make this sound like a person wrote it". The skill file loads into any agent that reads a SKILL.md. The author has not published a blind-judge test like other humanizer projects have. The only evidence right now is the before-and-after examples in the README.
Where this stops
An approved vocabulary of 900 words does not carry a brand voice. A marketing page written to the full standard reads as a technical manual. The skill is useful on drafts where the point is clear sentences, which covers most internal AI output: emails, support replies, change notes, sections of documentation. A product's own copy is still a writer's job.
Source
- GitHub, Dodaios/ste80-humanizer
- Hacker News, Show HN discussion
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.