humanizelisted
Install: claude install-skill sujan-mitta/ai-detect-humanize-skill
# AI Detect & Humanize
Two jobs, often done together:
1. **Detect** — estimate how AI-generated a text reads, and point at the exact passages that give it away.
2. **Humanize** — rewrite those passages (or the whole text) so they read like a thoughtful person wrote them, keeping the meaning, facts and format intact.
The scripts in `scripts/` do the mechanical parts (reading PDFs/DOCX, splitting huge texts, scoring, rebuilding the output file). You do the judgment and the writing.
## Be honest about what detection can do
No detector — commercial or this one — can prove who wrote a text. The scorer is a statistical model trained on labelled human and AI texts; the report prints its measured held-out accuracy, and it still misfires on some genres (polished human essays get flagged; carefully prompted AI slips through). Report results as "reads as likely AI / mixed / likely human", give the score, and say it's an estimate. Never tell a user a text is definitely AI-written or definitely human; people get accused unfairly on that basis.
Humanizing is about good writing: varied rhythm, concrete wording, a real voice. If a user says the goal is to submit AI work as their own where that's prohibited (e.g. a graded assignment), it's fine to mention once, briefly, that they should check their institution's rules — then help with the writing. Don't lecture or refuse.
## Step 0 — Get the text
| Input | Do this |
|---|---|
| Pasted text (short) | Use it directly. Save to a `.txt` f