The humanize Skill — Remove AI Fingerprints From Text
How the humanize skill scans for AI writing patterns across vocabulary, structure, formatting, and rhythm, then surgically removes them while keeping the author's message intact.
AI-generated text has a recognisable signature. The vocabulary clusters around the same words — "delve," "robust," "seamless," "leverage," "tapestry." The structure repeats: bullet points in sets of three, sentences that all open the same way, transitions that announce themselves. The humanize skill removes that signature without rewriting the content underneath it.
The skill takes a file path as its argument. Before making any change, it saves a timestamped backup next to the original and creates a working copy at a temporary path. No modification happens without the backup in place.
What the scan catches
The scan covers five categories. The first is banned vocabulary. English has a Tier 1 list of words that appear at high frequency in AI output and rarely in natural human writing: words like "delve," "meticulous," "pivotal," "leverage," "utilize," "robust," "seamless," "comprehensive," "empower," "elevate," "streamline," and several dozen more. Russian has its own equivalent list: overly formal filler phrases, textbook tone markers, and a set of English anglicisms that have exact Russian equivalents.
The second category is structural tells: over 40 percent of the text as bullet points, paragraphs that are all the same length, a formulaic section order (introduction, body, challenges, future outlook), and numbered lists that force ideas into arbitrary counts.
The third category is formatting artifacts: excessive em dashes, bold applied to every key term rather than used sparingly, stray markdown punctuation, and decorative horizontal rules between sections.
The fourth is communication artifacts: chatbot openers ("Certainly!", "Great question!"), knowledge-cutoff disclaimers, and acknowledgment loops. The fifth is linguistic patterns: copula avoidance ("serves as" instead of "is"), synonym cycling where the same thing gets three different names in three sentences, and transition overuse where "Moreover" and "Furthermore" appear paragraph after paragraph.
How fixes are applied
Changes go in three passes. The first handles vocabulary and formatting: Tier 1 words are replaced with natural alternatives, formatting artifacts are removed, chatbot artifacts are deleted. The second handles structure and rhythm: excessive bullet blocks are converted to prose, paragraph lengths are varied, synonym cycling is resolved by picking one term and keeping it. The third is an anti-AI audit: the edited text is re-read to check that no new patterns were introduced by the fixes themselves.
The core constraint is surgical editing, not rewriting. The author's structure, argument, and intent survive intact. Only the patterns that mark the text as AI-generated are removed.
After the scan
The skill shows a summary of changes by category — how many vocabulary replacements, how many structural fixes, how many formatting artifacts removed — and flags any change that altered meaning rather than style. Changes go to the original file only after confirmation.
Read what Datarim is for broader context, or see the related post on the /humanize command that uses this skill.