A Claude Skill that revises text and documents to remove the tell-tale signs of AI-generated writing, producing prose that reads as if a careful human wrote it.
Large-language-model output has statistical fingerprints: it smooths specific facts into generic, inflated, importance-signalling statements and reaches for the same small set of words, structures, and rhythms over and over. This skill reverses that — it restores specificity, cuts inflation, varies the rhythm, and prefers the plain word.
It works on pasted text or uploaded files (.txt, .md, .docx, .pdf).
Two deliberate guardrails keep the output good:
- It never invents facts to fill gaps. Removing puffery often means a sentence gets shorter or disappears — that's correct, rather than padding with made-up detail.
- It doesn't overcorrect into a new tell. Banning every em dash or making prose robotically terse is itself detectable. The aim is the prose a thoughtful human editor would produce, not the mechanical inverse of an LLM.
The full checklist lives in references/signs-checklist.md, organized into eight categories:
- Content & rhetoric — significance/legacy puffery, trailing "-ing" analysis, promotional tone, vague attributions, formulaic "Challenges/Future Prospects" sections.
- Vocabulary — the high-density "AI vocabulary" (delve, underscore, showcase, testament, tapestry, pivotal, robust, leverage…) and inflated synonyms (utilize→use, authored→wrote).
- Syntax — copula avoidance ("serves as" → "is"), negative parallelisms ("not just X, but Y"), the rule of three, forced non-repetition, knowledge-cutoff disclaimers.
- Structure & formatting — Title Case headings, excess boldface, bolded-inline-header lists, em-dash overuse, curly quotes, formulaic openings/closings.
- Markup & citation artifacts — Markdown bleeding into wikitext, leftover citation-bug tokens,
utm_source=chatgpt.com, placeholder text, hallucinated templates/citations. - Comment / talk-page / correspondence tells — canned good-faith boilerplate, exhaustive edit summaries, emoji formatting, wikilawyering moves.
- Historical (older-model) tells — didactic disclaimers, "In summary" sections, prompt refusals.
- What NOT to "fix" — weak signals where over-correction creates new problems.
| File | Purpose |
|---|---|
SKILL.md |
The skill itself — core principle, workflow, and the highest-signal fixes. |
references/signs-checklist.md |
The full sign-by-sign checklist with BEFORE → AFTER examples. |
de-ai-writing.skill |
Packaged, installable bundle of the above. |
Download de-ai-writing.skill and upload it in Claude's skill/capabilities settings, or point Claude Code / Cowork at this repository.
The checklist is distilled from Wikipedia's community field guide Wikipedia:Signs of AI writing, text of which is available under CC BY-SA 4.0.