Remove AI marks

Two kinds of trace survive in a file that a model helped produce. One is mechanical: characters and metadata a person never typed and cannot see. The other is the writing itself. This skill handles the first kind only.

Most of it is genuinely invisible: a zero-width space or an oaicite token changes nothing a reader would notice. One part is not, and it is worth being straight about it. Chatbot residue ("Certainly! Here's...", "أتمنى أن يكون هذا مفيدًا", "As of my last update...") is wording, and removing it is a small editorial act. Two rules keep it safe: a phrase is only removed when it is a whole sentence of its own, never a fragment of a longer one, and every removal is reported with the exact text that came out. So a line like "Please let me know if you have any questions" is left alone and flagged, because the sentence belongs to the writer even though it shares words with a sign-off.

The writing is a humanizer skill's job (arabic-humanizer for Arabic text). Keep the two apart. Mixing them produces the worst outcome, a file that was quietly reworded when the user only asked to clean it.

What it removes

Two deliberate exceptions, both about Arabic:

• Bidirectional marks (RLM, LRM, isolates) are reported, never removed. In a line like والـ ‏CTR‏ طلع 3.2% they are what keeps the Latin and the digits in the right place. Stripping them scrambles the line in a way nobody notices until a client opens the file.
• ZWNJ and ZWJ are removed only when the text has no Arabic in it. In Arabic, Persian and Urdu they control how letters join, so they are legitimate typography, not a hidden mark.

Running it

The work is done by scripts/clean_marks.py. It needs nothing but Python 3.7 and the standard library for text, Office files, PNG, JPEG and WebP. PDF uses pikepdf if importable, otherwise the qpdf command; if neither is there, the report says the PDF was not cleaned rather than implying it was.

python3 scripts/clean_marks.py inspect FILE [FILE ...]
python3 scripts/clean_marks.py clean   FILE [FILE ...] [--in-place]

When another skill called this one, run inspect only. A humanizer skill delivers its result in chat and writes the file once at the end, so producing a .cleaned file at that point hands the user a version they never approved. Report the findings, and let the caller fold the fixes into the text it is already showing.

When the user came here directly, inspect first and show what turned up, then clean. Two reasons this order is worth the extra second: the user gets to see that a client deliverable carried Author: ChatGPT, which is usually the thing they actually wanted to know; and a file with nothing in it needs no new version.

Options worth knowing:

Exit codes: 0 clean, 1 findings or changes, 3 something could not be fully processed. A 3 is not a pass; say which file and why.

One false positive to expect, because it will happen: a file that writes about these markers, a report, a log, a checklist, this skill's own text, gets flagged for naming oaicite or "As of my last update". The script matches strings and cannot tell a token from a mention of one. Read the finding before acting on it, and never clean a document whose subject is AI provenance without checking what the hits actually are.

Inside an Office file, forensic sentences are reported, not removed. A leftover "I hope this helps!" on a slide or a [Your Name] in a Word paragraph is named by inspect and left where it is by clean, which touches only characters in that file: deleting a sentence from a slide changes the layout, and a placeholder needs a real value. Those belong in the review loop a humanizer skill runs, not in a silent script pass.

Several files at once is normal, and a whole folder is a glob:

python3 scripts/clean_marks.py clean ~/Deliverables/*.docx ~/Deliverables/*.png

Reporting

State what was removed with counts, because a count is checkable and "cleaned the file" is not. Then state what was left and why: a placeholder that needs a value, a bidi mark that belongs there, a look-alike letter awaiting a decision. Name the output file.

What not to say: that the file is now undetectable, or that it will pass any particular checker. What this skill removes is real and verifiable. Whether some tool still flags the file is a different question and not one to promise on. A statistical watermark in the token choices, C2PA soft binding, and audio or video watermarks are all out of scope; say so if the user asks about them.

Then what

When the file holds prose the user is about to send, the mechanical pass is half the job. Say so in one line and offer a humanizer skill (arabic-humanizer for Arabic), which covers the writing: the staged openings, the "not X but Y" construction, the triads, the flat sentence rhythm. Do not start rewriting here.

When the request was to humanize in the first place and you arrived through a humanizer skill, run this pass, report the counts, and hand control back.

Ethics

This is hygiene for the user's own work: their metadata, their client files. Fine to clean a deliverable before it ships, or to keep an author field out of a document going to a client.

Not fine, and worth saying plainly once rather than helping around it: passing this off as proof a human wrote something, or defeating an academic integrity process or a disclosure a user is obliged to make. Clean the file, and let the claim about who wrote it be the user's own honest one.

Made by Hend Ibrahim · hendebrahim.com · More skills: github.com/hendibrahim-ai/claude-skills