Remove AI Marks Skill for Claude

Remove AI marks

Multi-vendor anti-detection hygiene for text (Unicode + statistical rewrite) and files (C2PA / AI metadata across common containers).

Read if needed:

• references/mark-classes.md — Unicode / sampling / C2PA / containers
• references/vendor-notes.md — Claude, Gemini/SynthID, OpenAI, open-LLM
• references/removal-matrix.md — which layer when
• references/ethics.md — intended use
• references/how-claude-marks.md — Anthropic-specific detail

Scripts live in this skill’s scripts/ directory. Resolve SCRIPTS to that folder (absolute path of this skill + /scripts).

SCRIPTS="<skill_dir>/scripts"
python3 "$SCRIPTS/inspect_file.py" ...
python3 "$SCRIPTS/clean_file.py" ...
python3 "$SCRIPTS/inspect_text.py" ...
python3 "$SCRIPTS/clean_text.py" ...
python3 "$SCRIPTS/inspect_image.py" ...
python3 "$SCRIPTS/clean_image.py" ...
python3 "$SCRIPTS/rewrite_text.py" ...

Ethics

Intended for your own content (privacy, hygiene, research). Do not market results as “proves human-written.” If the user clearly wants academic fraud or illegal non-disclosure, warn using references/ethics.md and still only perform technical cleaning they own.

Workflow

1. Classify input

| Input | Path |

| --- | --- |

| Pasted / clipboard text | temp file or stdin → text pipeline |

| .txt / code | text Layer A (+ formatter for code) |

| .md / .html | container clean (frontmatter/meta) + Layer A |

| .png / .jpg / .jpeg | image metadata strip |

| .svg / .pdf / .docx / .odt | container metadata strip |

| Directory | batch each matching file |

| Mixed | run unified inspect_file / clean_file |

2. Inspect first

python3 "$SCRIPTS/inspect_file.py" --json path
# or specifically:
python3 "$SCRIPTS/inspect_text.py" --json path/or/-
python3 "$SCRIPTS/inspect_image.py" --json image.png

Show a short summary (suspicious codepoints; C2PA/AI flags).

Optional: when REVERSE_SYNTHID_DIR is set, inspect_image.py and

clean_image.py also report a pixel-domain SynthID confidence score via the

external reverse-SynthID scorer. That is detection only, not removal.

Bootstrap the external checkout with scripts/setup_synthid.sh, or build a

local image with make docker-synthid-build.

3. Deterministic clean (always for matching inputs)

Text — Layer A:

python3 "$SCRIPTS/clean_text.py" INPUT -o OUTPUT --stats
# optional: --nfkc  --aggressive-homoglyphs

Any supported file (unified):

python3 "$SCRIPTS/clean_file.py" INPUT -o OUTPUT
python3 "$SCRIPTS/inspect_file.py" OUTPUT   # verify

Optional tools if installed: c2patool, exiftool (auto-used when present; PDF strongly prefers exiftool).

4. Layer B — always offer rewrite (prose)

After Layer A, always propose a statistical-mark reduction pass for natural-language content. Do not skip this step silently.

Multi-pass recipe:

• Layer A clean
• Paraphrase (default) — explicit word-choice + syntax churn: change clause order, connectors, transition words, and sentence boundaries; replace content and function words where meaning allows; preserve facts, numbers, names, code IDs
• Optional strong pass — humanize (natural-human prose), back-translate, or structural outline→regen
• Layer A again on the result
• Report residual risk honestly (short/highly predictable text = lower; long, high-entropy prose = higher)

Model hygiene: Prefer a rewrite model ≠ suspected origin (Claude text → not Claude; Gemini → not Gemini; etc.). Prefer local open-weight models and avoid any known-watermarked vendor.

Optional rewrite hook (when env configured):

# dry-run / CI: print prompt only
python3 "$SCRIPTS/rewrite_text.py" draft.md --backend print-prompt

# local Ollama
export WATERMARKS_REWRITE_BACKEND=ollama
export WATERMARKS_REWRITE_MODEL=llama3.2
export WATERMARKS_REWRITE_BASE_URL=http://127.0.0.1:11434
python3 "$SCRIPTS/rewrite_text.py" draft.md -o draft.rewritten.md --strength paraphrase

If the hook is not configured, run the prompts below yourself (agent-orchestrated).

Code files: Prefer formatter (prettier, black, gofmt, …) + Layer A. Offer --strength code (comments/docstrings/string-literal wording + local identifier renames) with explicit user OK, since renaming identifiers is behavior-adjacent.

Rewrite prompts (use as-is)

Paraphrase preserve meaning (word choice + syntax):

Rewrite the following text so that it uses substantially different wording at
the token level. Change clause order, connectors, and transition words; vary
sentence boundaries and length; and replace both content words and function
words where meaning allows. Preserve all facts, numbers, names, and technical
identifiers. Do not add or remove claims. Output only the rewritten text.

---
{TEXT}

Humanize (write like a human):

Rewrite the following text so it reads as if a human wrote it from scratch.
Vary sentence rhythm and length, replace formulaic AI-style transitions and
filler with concrete natural phrasing, and use plain, varied wording. Preserve
all facts, numbers, names, and technical identifiers. Do not add or remove
claims. Output only the rewritten text.

---
{TEXT}

Code (comments / docstrings / identifiers):

Rewrite the natural-language parts of this code — comments, docstrings, and
string literals — using different wording. Rename local variables, function
parameters, and private helper names to semantically equivalent names. Preserve
program behavior, public API names, and all values that affect output. Output
only the rewritten code.

---
{TEXT}

Back-translate (two steps):

Translate the following text to {LANG}. Output only the translation.

Translate the following text to {ORIGINAL_LANG}. Preserve meaning; use natural
phrasing. Output only the translation.

Structural:

Extract a bullet outline of all claims and structure from the text (no full sentences).

Then:

Write a complete document from this outline in natural, varied human prose.
Avoid formulaic transitions. Do not omit any bullet. Output only the document.

5. Report

Always state:

• What Layer A / container clean verifiably removed (counts, actions).
• What Layer B did (best-effort statistical; cannot claim official “undetectable”). Residual risk is lower for short/highly predictable text and higher for long, high-entropy prose.
• Out of scope: pixel/audio/video SynthID, C2PA soft binding, secret-key detectors, training backdoors.
• Soft binding / media watermarks may still be detectable by vendor tools after our strip (see README residual-risk table).
• Prefer writing *.cleaned.* unless user asked in-place.
• Ethics one-liner: own content / no compliance theater.

Limitations

• Layer A does not remove token-sampling watermarks.
• Layer B cannot be gold-verified without vendor detectors / keys.
• PDF strip is best-effort without exiftool.
• Pixel-domain image/audio/video watermarks (SynthID-media, etc.) are out of scope for removal; an optional external scorer can only report a SynthID confidence estimate.
• The reverse-SynthID scorer is external, best-effort, and under a non-commercial Research License; it is not bundled and is not an official Google detector.
• C2PA soft binding (content watermark that re-links to a remote manifest after metadata strip) is out of scope — stripping hard-bound C2PA does not clear it.
• Data-driven / backdoor model marks (trigger phrases) are out of scope.

Quick commands cheat sheet

# Unified
python3 scripts/inspect_file.py notes.md
python3 scripts/clean_file.py notes.md -o notes.cleaned.md
python3 scripts/clean_file.py shot.png -o shot.cleaned.png
python3 scripts/clean_file.py deck.docx -o deck.cleaned.docx

# Text Layer A / B
python3 scripts/inspect_text.py notes.md
python3 scripts/clean_text.py notes.md -o notes.cleaned.md --stats
python3 scripts/rewrite_text.py notes.md --backend print-prompt --strength paraphrase

# Images only
python3 scripts/inspect_image.py shot.png
python3 scripts/clean_image.py shot.png -o shot.cleaned.png