| name | launder |
| description | Wash the mechanical fingerprints out of text — zero-width and invisible characters, smart quotes, em/en dashes, the unicode ellipsis, non-breaking and exotic spaces, soft hyphens — so the bytes read like a human typed them. It scrubs typography, it does NOT rewrite prose (that's the model's job) or detect word-level tells (that's `tell`'s). Use before pasting model output where invisible characters or fancy punctuation would give it away, corrupt a diff, or break a downstream parser. |
launder
wash out the prints.
A model doesn't just pick suspicious words — it leaves suspicious bytes.
Zero-width spaces sprinkled mid-word, curly quotes, em-dashes, that one
non-breaking space, the unicode ellipsis. None of it is visible; all of it
marks the text as machine-produced (and some of it is literally watermarking).
This skill is for washing those prints off — touching the typography, never the
words.
launder is the cleanup sibling of tell. tell detects the AI smell;
launder washes out the invisible and typographic giveaways. tell reads the
prose and names the overused words; launder scrubs the bytes underneath it.
Rules
- Scrub the invisibles. Strip zero-width spaces, word joiners, the BOM,
and soft hyphens. They carry no glyph — only a fingerprint. Delete them.
- Straighten the punctuation. Curly quotes → straight, em-dash →
--,
en-dash → -, unicode ellipsis → ..., non-breaking and exotic spaces →
a plain ASCII space. Normalize the typography to what a keyboard produces.
- Don't touch the words. Re-phrasing, de-clichéing, killing "delve" — that
is not laundering, that is rewriting, and it is not this skill's job. launder
changes bytes, never meaning.
- Leave clean text alone. Plain ASCII passes through byte-for-byte. If
there are no prints to wash, do nothing and say so.
- Homoglyphs are opt-in. Mapping Cyrillic/Greek look-alikes (
а→a,
о→o) back to ASCII is lossy — a genuinely Cyrillic word would be
mangled — so only do it when explicitly asked (--homoglyphs).
What launder is NOT
- It is not a paraphraser. It will not reword a sentence, vary your
vocabulary, or restructure a paragraph. It only normalizes characters.
- It is not a
tell replacement. tell finds the word-level tells;
launder finds the byte-level ones. Run tell to know why prose reads as
AI; run launder to remove the typographic evidence.
- It is not a cloak. Honest note: stripping zero-width characters and
smart quotes removes formatting artifacts — it does not make machine text
"undetectable" and is not a way to defeat plagiarism or AI-detection checks.
The words still read like the model wrote them. launder cleans bytes; it does
not launder authorship.
Example
Before: he said “hi” — with curly quotes and a zero-width word joiner
after the closing quote (invisible here, but present in the bytes).
After: he said "hi" — straight quotes, the invisible character gone,
nothing else changed.
Output
Deliver the laundered text, not a report on laundering it. Don't narrate running
launder.py. If the text was already clean, say so in one line; if asked what
was found, give the category list plainly. No preamble, no closing note.
Companion tool
launder.py in this folder applies these scrubbers mechanically. Pipe text
through it to clean it (launder.py), gate a pipeline (launder.py --check,
exits 1 if any fingerprint is present), list what it found by category
(launder.py --report), or get a structured summary (launder.py --json).
Add --homoglyphs to also normalize confusables. It pairs with tell:
tell diagnoses the AI prose, launder washes the AI typography.
Extend the scrub table with --patterns FILE (repeatable; or the
LAUNDER_PATTERNS env var). The file is JSON of shape
{"<category>": {"<char>": "<replacement>"}}, e.g.
{"smart_quote": {"«": "\""}, "bullet": {"•": "-"}}. Known categories extend
the built-ins, new categories are allowed and reported under their own name, and
user entries override the built-ins on character collision.
