Voice Restore — AI skill for Claude Code

voice-restore

**Find out whether a text was AI-touched — then put the author's own voice back.**

Not "humanize". Humanizing optimises for *reads like a person*. This optimises for *reads like **this** person*. On real corpora those two goals conflict, and when they do, this skill sides with the author.

---

Why the obvious approach is wrong

Every "reduce AI detection" guide says the same thing: vary sentence length, add short sentences, insert personal voice. That advice assumes the target is a generic human.

It isn't. Take one measured case — a 7,766-word finance thesis flagged at 46%:

	the thesis (flagged 46%)	the same author's undetected earlier essay
sentence-length cv	0.518	0.286
article-initial sentences	35.9%	41.0%
short sentences (≤10w)	10%	6%

The flagged document is **more varied** than the author's own writing, and uses **fewer** article-initial sentences. By the standard burstiness playbook, it was already "more human" than he is.

So the playbook would have made it worse. What it needed was the opposite: flatten it toward *his* distribution.

⛔ Shape metrics measure **how smoothed a text is**, not **whether a model wrote it**. A writer with imperfect English is flat by default. Flat is his fingerprint, not his crime.

---

What actually survives AI editing

Three things, and they are cheap to count:

Tell	What it looks like	Why it's hard to explain away
First person = 0	7,766 words, not one I / we / us / our / my	Nobody writes their own thesis without once saying remind us
Prose semicolons appear	author's own writing: 0 in 3,371 words; polished draft: 16	Semicolons weld two independent sentences together — that is an editing move, not a drafting move
The up, This down	author: The 14.1% / This 19.2%; draft: The 25.8% / This 6.2%	He delegates anaphora to This. The polish pass rewrote every one as