Built on Andrej Karpathy's October 2, 2026 suggestion to prompt models "80% of the way to ASD-STE100 Simplified Technical English," the MIT-licensed skill strips AI tells and applies controlled-language structure rules — and is explicit that beating AI detectors is not its goal.
2026-11-10 · by Moe Ameen
Dodaios, the GitHub organization behind an open-source "agent OS" project (formerly called Personal Jarvis), released STE80 Humanizer, an open-source "agent skill" that rewrites AI-generated text so it reads more like a person wrote it — without changing the meaning. It is distributed under the MIT license and installs into coding agents rather than running as a website: it works with Claude Code, Codex, Cursor, and Claude.ai/Desktop, and ships a small dependency-free Python linter (`scripts/ste_check.py`) alongside the skill files.
The project implements an idea Andrej Karpathy floated in a post on October 2, 2026: prompt a language model to explain things in ASD-STE100 Simplified Technical English, the controlled-English standard originally written for aircraft maintenance manuals, because models know the standard well and its "heavy constraints on clean writing style" produce text he found more readable. Because the full standard is strict, Karpathy softened the request to roughly "80% of the way to ASD-STE100" — which is where the skill gets its name. ASD-STE100 is maintained by the ASD Simplified Technical English Maintenance Group; it pairs a set of writing rules with a controlled dictionary of about 900 approved words, has been free of charge since Issue 6 in 2013, and reached Issue 9 in January 2025.
STE80 Humanizer runs in two passes. Pass 1 removes a set of 26 "AI tells" drawn from Wikipedia's "Signs of AI writing" and the catalog in Siqi Chen's MIT-licensed blader/humanizer project. Pass 2 applies 15 STE-style structure rules — one idea per sentence, short sentences, active voice, and similar. Three modes tune the aggressiveness: Humanize (the default), Light (Pass 1 plus only the worst Pass 2 offenders), and Strict STE (all of Pass 2). It can also match a voice from a user-supplied writing sample and runs a "fidelity" focus that keeps facts intact and flags contradictions in the source.
The author is careful about scope. The skill paraphrases the ASD-STE100 rules and does not bundle the specification or its dictionary; it is not affiliated with or endorsed by ASD, and the README states plainly that "STE style" is not "STE compliant," which requires the official dictionary, trained writers, and formal review. Most pointedly for the current moment, the project says getting past AI detectors is explicitly not a goal — the aim is readability. Its linter is described as a heuristic, "not an AI detector and not a compliance checker." Treat the exact rule counts and mode behavior as a snapshot and confirm them against the repository, since an actively developed skill changes.
The lesson worth taking from STE80 Humanizer is not "run a linter after you write." It is that a controlled, readable style is something you can specify up front — and the cheapest place to enforce it is at generation, not in a cleanup pass on finished text. That is the distinction [Kompozy](/) is built around. Kompozy is a content generation and publishing engine, and when it drafts Text Posts, a Blog Article, or an Email Newsletter, the [Persona Brief](/glossary/persona-brief) fixes the voice and a banned-word filter strips the generic AI tells — the flat openers, the rule-of-three padding, the reflexive "it's not X, it's Y" — that STE80's Pass 1 is designed to remove. You can hand that brief the same instinct Karpathy described ("write about 80% of the way to ASD-STE100: one idea per sentence, short sentences, active voice") so clarity is a property of every draft instead of a chore you remember to do.
Then Kompozy carries that clean copy past the one thing a text-only skill can't touch: format and distribution. The same governed idea fans out into [Carousel Posts](/glossary/hyperframes), quote graphics, a [Persona Short](/glossary/persona-shorts) narrated by a recurring avatar, and a newsletter — and [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email, behind a per-post review gate where a human still approves every piece. STE80 makes one block of text clearer in your editor; Kompozy makes clarity the default across a week of multi-format output and ships it. If detector anxiety is what pushed you toward a humanizer in the first place, the honest posture — generate in a real voice, review it yourself, disclose how you work — is the one STE80 and Kompozy both point at, and the one that outlasts any single scanner.
STE80 Humanizer is a free, open-source agent skill (MIT license) published by Dodaios (a GitHub organization) that rewrites AI-generated text to read more naturally without changing its meaning. It installs into coding agents — Claude Code, Codex, Cursor, and Claude.ai/Desktop — and includes a small dependency-free Python linter. It works in two passes: one that removes common "AI tells," and one that applies ASD-STE100-inspired structure rules like one idea per sentence and active voice.
No, and the project says so explicitly. Unlike tools marketed to evade Turnitin or GPTZero, STE80 states that getting past AI detectors is not a goal — the aim is readability. Its linter is described as a heuristic, not an AI detector and not a compliance checker.
ASD-STE100 Simplified Technical English is a controlled-English standard created for aerospace maintenance manuals, maintained by the ASD Simplified Technical English Maintenance Group. It combines writing rules with a dictionary of about 900 approved words and reached Issue 9 in January 2025. The "80%" comes from a suggestion Andrej Karpathy made on October 2, 2026: because the full standard is strict, he softened the ask to writing "about 80% of the way to ASD-STE100," which keeps the clarity with a more natural flow.
Those are detector-evasion humanizers built to make AI text score as "human" on detectors. STE80 is the opposite framing — it optimizes a reader's comprehension and openly disclaims detector evasion. And all three are text-rewriters: none generates new content, enforces a persistent brand voice across formats, or publishes. A content engine like Kompozy governs voice at generation time with a Persona Brief and banned-word filter, then produces and schedules video, images, carousels, blog, and newsletter across platforms.