ste80-humanizer: STE80 Humanizer: agent skill that removes AI tells and rewrites text 80% of the way to ASD-STE100 (Simplified Technical English), the method Andrej Karpathy shared
Core Highlights
Dodaios/ste80-humanizer is an agent skill built to strip the "AI smell" out of machine-generated text and rewrite it roughly eighty percent of the way toward ASD-STE100, a simplified technical English standard. ASD-STE100 is the controlled-language specification long used for aviation maintenance documentation, prized for its fixed vocabulary, short sentences and near-total absence of ambiguity; Andrej Karpathy publicly recommended the approach as a way to improve technical writing. The project has around 696 GitHub stars, is written in Python and carries the MIT license. Put simply, it targets a problem that keeps growing: AI prose is instantly recognizable as AI, and this skill makes it read more like something a person — specifically a professional — would write.
What It Does
The skill works by recognition and rewriting. It first identifies the common tells of AI text — excessive connectives, vague adjectives, neat parallel structures, stock transitions — then rewrites them according to simplified technical English rules: short sentences, active voice, concrete nouns and a restricted vocabulary. It is not a standalone app but a skill file that coding agents such as Claude Code, Cursor and Codex can load; once installed, the agent follows the writing conventions whenever it produces documentation. The emphasis on "80 percent" rather than one hundred is deliberate: full ASD-STE100 conformance requires a certified word list and human review, which is overkill for everyday writing.
Technical Details
The heart of ASD-STE100 is a controlled vocabulary plus a set of writing rules: each word has a single meaning, nominalizations are avoided, one sentence expresses one idea, and active voice is preferred. ste80-humanizer compresses these rules into instructions an agent can read, so the model sidesteps AI mannerisms at generation time. The implementation is deliberately light — essentially a structured rules/prompt file paired with a skills.sh installation entry. A validate GitHub Action checks the rule file for formatting and consistency. Because it is a skill rather than a particular model, it can be reused across tools and across models.
Comparison With Alternatives
Most "humanize AI text" tools are rewriters or detectors that pass text through another model for a polish pass. ste80-humanizer takes a different route: it moves the convention upstream, so the agent writes in controlled language from the start. That is less work and more consistent than after-the-fact rewriting. Compared with generic humanizing prompts, it rests on the mature ASD-STE100 standard, which makes its rules concrete and verifiable rather than a matter of taste.
Industry Impact and Use Cases
For technical writers and people in aviation or engineering documentation, it offers an automatable path to clear writing. For general users, it is a practical plugin that makes AI output sound more natural and human. The limitations are real: controlled language produces a plain style unsuited to creative writing, and targeting eighty percent means a few traces survive and still need a human pass.
Practical Notes
The skill fits best where clarity matters more than flair — API docs, internal runbooks, support articles and compliance text. Teams can pair it with a house style guide, letting the skill enforce sentence-level discipline while humans handle structure and tone. Because it operates as prompt-level rules, results vary somewhat by model: stronger instruction-following models honor the constraints more faithfully. Users should also keep the validate action in their fork so rule-file edits do not silently break the skill.
Why It Resonates Now
The timing of this project is no accident. As AI-generated documentation, README files and support articles flood code repositories, readers have grown adept at spotting the telltale rhythms of machine prose, and that recognition erodes trust. ste80-humanizer answers by making the output rules explicit and auditable rather than relying on a model's vague notion of "writing naturally." The aviation roots of ASD-STE100 are a deliberate choice: that standard was built for situations where a misunderstood sentence can have costly consequences, so its bias toward one-idea sentences and unambiguous words translates well to technical documentation of any kind. It is worth noting that the skill changes style, not substance — it will not fix a wrong claim or a missing fact, only present correct information more cleanly. Users who pair it with a solid editorial pass get the best of both: machine speed on phrasing, human judgment on accuracy. As agent skills become a common way to distribute reusable behavior, projects like this hint at a broader shift, where writing standards are packaged as installable modules instead of style guides that nobody reads.