Dodaios/ste80-humanizer

What it does

STE80 Humanizer is an agent skill for rewriting AI-sounding prose. It runs two passes: first it removes the 26 "signs of AI writing" catalogued by Wikipedia and blader/humanizer, then it rebuilds sentences toward 80% of ASD-STE100, the controlled English used in aircraft maintenance documentation. The 80% target, which Andrej Karpathy described as more readable than strict STE, keeps normal vocabulary while enforcing one idea per sentence, active voice, and simple tenses instead of the full controlled dictionary.

The output is aimed at readers who want plain explanatory text, not at passing AI detectors. The README is explicit that evasion is not a goal. It suits documentation, runbooks, READMEs, and emails where clarity matters; fiction, speeches, and brand copy get only the first pass because the method targets explanatory writing.

The skill encodes rules as numbered checks. SKILL.md lists 26 tells as §1–§26 and 15 STE80 structure rules as S1–S15, and the agent reports edits by number so you can look up what changed. Three modes control depth: the default humanize mode, a light touch, and strict STE.

Quick start

The easiest route is the Vercel skills CLI, which installs the skill for Claude Code, Codex, Cursor, and other agents:

npx skills add Dodaios/ste80-humanizer

Add --global --agent claude-code (or codex, cursor, or '*' for every agent) to install outside the current project. Claude Code users can instead add it as a plugin through /plugin marketplace add Dodaios/ste80-humanizer and /plugin install ste80-humanizer@ste80-humanizer; Claude.ai and Claude Desktop users upload the repository ZIP under Settings → Capabilities → Skills install options.

After installation, invoke it with /ste80-humanizer and paste the text, or ask in plain language, e.g. “Rewrite this README 80% of the way to ASD-STE100” or “Convert docs/runbook.md to strict STE.” You can also check prose with the bundled dependency-free linter:

python scripts/ste_check.py README.md
python scripts/ste_check.py --strict docs/runbook.md

The checker flags both AI tells and measurable STE rules such as sentence length, semicolons, passive voice, and "-ing" forms, but the author calls it a heuristic, not a detector.

Constraints

• STE is an English standard. For other languages the skill applies the tell list and structural rules but not the English word lists.
• Strict mode is not certified compliance: real ASD-STE100 conformance requires the official dictionary and trained review.
• The repository does not include the ASD-STE100 specification or dictionary; rules are paraphrased, and the project states it is unaffiliated with ASD.
• The skill works through supported agent hosts, so output quality depends on the underlying model and the host's skill support.
• Code is MIT-licensed, and the tell catalogue derives from the MIT-licensed blader/humanizer.

HotRepo's take

The repository entered HotRepo's new-project list on 2026-11-10 with 103 stars and 6 forks gained since 2026-11-09, and its Show HN thread drew 59 points and 16 comments. Discussion frames the project less as a humanizer than as a clarity tool: one commenter argues it removes AI tics without adding human voice thread, while another notes that simply prompting for "80% of the way to ASD-STE100" already improved readability noticeably comment. Readers comparing it with similar STE skills will find the two-pass pairing and the numbered, auditable edit rules to be the distinguishing detail, and the author's own note that gains showed up in factual fidelity rather than style on named-STE prompts is a useful calibration of what the skill adds.