| name | scrub |
| description | Local offline hygiene for media and documents the user owns. Strips EXIF/IPTC/XMP, best-effort C2PA/JUMBF, hidden Unicode (zero-width, bidi, tag chars), and document properties in PDF/DOCX/ODT/HTML/SVG/Markdown; optional spectral disruption for images/video/audio. Use when the user asks to scrub, clean metadata, strip C2PA, Content Credentials, EXIF, hidden Unicode, AI marks, SynthID-class research transforms, or inspect provenance on files they have rights to modify. |
Scrub
Run the repo CLI. Do not upload files. Do not claim a detector will fail.
Rules
- Only process files the user owns or has explicit rights to modify.
- Originals stay untouched; outputs use a
_clean suffix.
- Say best-effort / degrade / strip tags — never “guaranteed bypass” or “undetectable.”
- Statistical (token-sampling) text watermarks are out of scope. Unicode hygiene is not a paraphrase rewrite.
Commands
From the Scrub repo root (prefer .venv/bin/python if present):
python cli.py inspect PATH
python cli.py clean PATH -o OUTPUT_DIR
python cli.py clean PATH -o OUTPUT_DIR --metadata-only
| Input | What clean does |
|---|
.jpg .png .webp .heic … | Metadata + optional spectral/spatial chain |
.mp4 .mov .wav .mp3 … | ffmpeg remux/re-encode, drop tags; optional spectral on short video |
.md .txt .html .svg | Hidden Unicode + obvious generator/C2PA-ish meta |
.docx .odt | Core/app props + XML text hygiene |
.pdf | Info/XMP via pypdf or exiftool when installed |
Workflow
- Confirm the path and that the user wants a copy, not in-place overwrite.
inspect first if they asked what is in the file.clean with --metadata-only unless they asked for image/video signal disruption.- Show the output path and quote the CLI
detail line. Offer NOTICE.md if they ask about legality.
Install (optional, user machine)
mkdir -p ~/.cursor/skills && ln -sfn "$(pwd)/skills/scrub" ~/.cursor/skills/scrub
mkdir -p ~/.claude/skills && ln -sfn "$(pwd)/skills/scrub" ~/.claude/skills/scrub
