Image Metadata Cleaner

AIwork4me/image-metadata-cleaner

image-metadata-cleaner cleans privacy-sensitive metadata from user-owned image files by writing separate sanitized copies. This skill is for privacy hygiene and reproducible file preparation, not for hiding authorship, evading provenance checks, bypassing platform labels, or misrepresenting where an image came from.

Safety boundary

• Use this skill only for images the user owns or is authorized to process.
• Do not run it for requests whose stated purpose is to bypass AI labels,

defeat detection systems, fake provenance, or violate a platform's disclosure rules. In those cases, explain that you can help with privacy-preserving metadata cleanup, but not deception or evasion.

• Never edit the original image in place. The script refuses an output path that

resolves to the same file as the input.

• Do not use --overwrite unless the user explicitly asks to replace existing

output copies.

Before running

Confirm the input path, output location, and output format unless the user explicitly invoked this skill with those arguments. If the request is a folder operation, prefer the default metadata-cleaned/ output folder so generated copies are separated from source files.

Run a dry run when the user asks to preview changes:

uv run --with pillow==12.2.0 python "${CLAUDE_SKILL_DIR}/scripts/strip.py" "" --dry-run

Execution

Single file, safe default output name beside the input:

uv run --with pillow==12.2.0 python "${CLAUDE_SKILL_DIR}/scripts/strip.py" "" --manifest

Single file with an explicit output file:

uv run --with pillow==12.2.0 python "${CLAUDE_SKILL_DIR}/scripts/strip.py" "" --output "" --manifest

Folder batch, writing copies into /metadata-cleaned/:

uv run --with pillow==12.2.0 python "${CLAUDE_SKILL_DIR}/scripts/strip.py" "" --manifest

Recursive folder batch with a separate output directory:

uv run --with pillow==12.2.0 python "${CLAUDE_SKILL_DIR}/scripts/strip.py" "" --recursive --output-dir "" --manifest

Output format defaults to preserve: JPEG inputs stay JPEG, and other supported formats are written as PNG to avoid unexpected lossy conversion. If the user provides an explicit --output path ending in .jpg, .jpeg, or .png, the script follows that extension unless --format is set. Use --format jpg only when the user explicitly wants JPEG output; transparent images will be composited onto a white background.

After running

Report the script results without overstating them:

1. Number of files processed, previewed in dry-run mode, or failed.
2. Output filenames and where the manifest was written.
3. File size before and after.
4. Dimensions reported after re-encoding.
5. Whether the verification scan found common metadata keys or provenance marker

strings in the output.

Do not claim that all possible watermarks or provenance signals were removed. The script removes file-level metadata visible to Pillow-style re-encoding and performs a basic marker scan; it does not remove pixel-level watermarks, fingerprints, or external platform records.

Supported inputs

.png, .jpg, .jpeg, .webp, .bmp, .tiff, .tif

Troubleshooting

• No supported image files found: The folder contains no files with supported

image extensions, or all candidates are inside the output directory.

• Output already exists: The user supplied an explicit output path that

exists. Choose another path or rerun with --overwrite only after user confirmation.

• Unsupported image / cannot identify image file: The file extension looks

supported, but Pillow could not decode the file. Report the file and continue with any other batch items.

• uv not available: Ask the user before falling back to a local Python

environment with Pillow installed.

Technical details

For implementation notes, verification limits, and metadata caveats, read references/technical-details.md.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AIwork4me
• Source: AIwork4me/image-metadata-cleaner
• License: MIT-0

Install and usage instructions live in the source repository linked above.