Claude Skill for Humanizing AI Content

Teach Claude to think in two passes: Pass one is diagnosis - Scan for banned phrases ("leverage," "robust," "delve") - Flag repetitive sentence lengths - Mark facts that must stay untouched (numbers, names, dates) - Identify your audience and intent Pass two is reconstruction - Rewrite with the diagnosis in mind - Preserve every fact you marked - Vary sentence rhythm naturally - Replace abstract verbs with concrete ones - Cut template structures Instead of giving rules, feed Claude reference materials it can read. Create a references folder with: - Rubric file (8 criteria for what "human enough" means) - Edit-library file (before/after examples of each transformation type) - Taboo-phrases file (categorized list of AI-isms to eliminate) - Fact-preservation file (rules for what never changes) Then build validation scripts: - One checks if facts got preserved - One scans for banned phrases - One measures readability metrics - One flags if you changed too much The Skill file itself just routes traffic... it reads your input, picks which preset to use, runs the two passes, validates output. AI humanization fails a lot because it's vague... "write like a human" means nothing to an LLM. But "score above 4 on natural rhythm, eliminate these 17 phrase patterns, keep sentence variance between 8-25 words" is executable.