A Claude Code skill that checks your writing against your own voice, not generic AI-pattern rules.
Started from Siqi Chen's blader/humanizer and diverged into a voice-first fork. Blader catches generic AI tells; this catches those too, plus a voice check that measures drift from your specific writing — a private corpus of your own work is the reference, not a universal list. The pattern lists have grown apart: some overlap, some patterns unique to each fork (including one from blader's PR queue that never landed upstream).
For drafts:
- Voice drift flags (with rewrites in your register)
- Specificity audit (every unanchored claim, with required substitution)
- AI pattern check (weighted below voice drift)
- Comment stacking detection (attribution lists that aren't building toward a point)
For incoming comments:
- Authenticity score (0–3): specific citation + personal experience + real question
- Synthetic praise detection
- Response length recommendation
For your own replies:
- Three-step engagement check: locate what's new → extend it → redirect
- Voice drift flags
- Tic detection
Here's voice-humanizer catching drift a generic humanizer would miss.
Before (draft):
Skills in Claude Code are stored across three locations. Understanding the sync behavior helps you avoid unexpected issues. The install directory holds runtime files, the manifest cache stores metadata, and your git repo is the source of truth.
Flag:
Voice drift — three-item list where your corpus compresses to two. "Understanding X helps you avoid Y" is the signpost-before-the-move construction (#33) — your corpus states the failure, not the promise. Parallel three-clause structure ("A holds B, C stores D, E is F") shows uniform investment (#36) where your corpus is uneven.
No AI pattern fires hard. Grammar's clean. But it doesn't sound like you.
After:
Skills live in three places: the install directory (runtime files), the manifest cache (metadata), the git repo (source of truth). Edit one and forget the others, next sync clobbers you. Ask me how I know.
A generic humanizer would pass the Before. Voice-humanizer caught it because the corpus knew this author compresses lists, states failures instead of announcing help, and closes on lived experience. That's the difference.
mkdir -p ~/.claude/skills
git clone https://github.com/dannwaneri/voice-humanizer.git ~/.claude/skills/voice-humanizerThen replace CORPUS.md with your own writing. See SETUP.md for full instructions.
Generic humanizers check for patterns to avoid. This one checks for patterns to match — your specific cadences, the kinds of specificity you reach for, what good writing in your voice actually sounds like vs what Claude-flavored-you sounds like.
The corpus is private (.gitignore). The skill and structure are public.
Built for Claude Code. Claude desktop users see SETUP.md for manual usage. Inspired by blader/humanizer.