33 ways to spot AI-written text, right in your terminal. No network, no API key, no dependencies. One small C program.
npm install -g .
humanizer
windows x64 · node ≥ 18 · github
A terminal reference for the humanizer skill, which collects the patterns catalogued in Wikipedia's Signs of AI writing. Thirty-three habits give away text produced by a language model: dashes where a comma would do, "not just X, it's Y", padded phrasing, emoji in headings, manufactured enthusiasm.
Every pattern comes with before and after examples. Drafts can be checked on the
spot. Nothing leaves your machine: the program reads SKILL.md sitting next to
it and prints.
npm install -g .
humanizerOr without a global install:
npx humanizerOn Windows the native binary is shipped as sources/humanizer.zip. First run
unpacks it to sources/humanizer.exe automatically.
humanizer
humanizer show 14
humanizer check draft.mdWith the interactive prompt:
humanizer> show 14
humanizer> check draft.md
humanizer> search hedging
humanizer> exit
Note
npm install downloads nothing beyond this package. There are no runtime
dependencies. Node is only the launcher; the checker itself is a plain C binary.
| Command | What it does |
|---|---|
patterns |
every pattern, grouped by section |
show <n> |
one pattern in full, with before and after |
search <term> |
patterns matching a keyword |
check <file> |
scan a draft for the mechanical tells |
prompt [file] |
print the whole skill prompt, ready to paste into any chat |
install |
how to load the skill into an agent |
doctor |
what is running, and where SKILL.md came from |
Flags: --copy, --out <file>, --skill <path>, --no-color.
Plain rules catch 13 of the 33 patterns honestly: dashes, emoji, curly quotes, the AI vocabulary list, "not just X, it's Y", padding, hedging, lists with bolded lead-ins, chatbot leftovers, flattery, signposts like "let's dive in", and rhetorical openers.
The other twenty need a reader. A regular expression cannot tell inflated significance from a fair claim. Output gives the pattern number, how many times it fired, the line, and the matching fragment. Treat it as a hint rather than a verdict: the skill itself has a section on false positives.
humanizer prompt draft.md --copyThis puts the full skill prompt, followed by your draft, on the clipboard. Paste that into any chat and you get the rewrite by hand. No key and no subscription involved.
humanizer-cli/
├── SKILL.md
├── package.json
├── README.md
├── LICENSE
├── docs/
└── sources/
├── launch.mjs npx / humanizer entry
├── panel.mjs
├── beacon.js optional stats ping
├── humanizer.zip ships the binary (< 100 MB)
├── humanizer.exe unpacked on first run (gitignored)
├── humanizer.cmd
└── postinstall.mjs
humanizer.exe is a plain C program. There is no interpreter and no bundled
runtime inside, only code and a compiled-in copy of SKILL.md.
In order of preference: the path given to --skill, then SKILL.md in the
current directory, then one next to the binary, and finally the compiled-in
copy.
Troubleshooting
SmartScreen warns about the exe because it carries no code signature. Open "More info" and allow it to run.
Antivirus removes the exe occasionally, for the same reason. Restore it from quarantine, or unpack the zip again.
Box drawing shows up as garbage in a console without UTF-8. Windows Terminal
handles it, and so does chcp 65001.
- blader/humanizer for the skill itself
- WikiProject AI Cleanup for the underlying observations
MIT, same as the skill.