Strips AI writing patterns from text and rewrites it to sound genuinely human — removing the statistical defaults, then adding earned voice calibrated to genre (opinion pieces get a person's voice; docs and summaries stay neutral) without ever faking humanity. Use when a draft reads as AI-generated, over-polished, or rhythmically uniform — including blog posts, emails, LinkedIn posts, or any prose that needs to sound like a real person wrote it. Produces a pattern audit, side-by-side comparison, itemised change log, and clean rewritten output ready to paste.
“Humanize this text: [paste]”“Use the notes-humanizer skill on this draft”“This reads like ChatGPT wrote it — fix it: [paste]”“Strip the AI out of this and make it sound like a real person wrote it”“Run the humanizer on this LinkedIn post: [paste]”
What to give it
▸Text to humanize — Any length. Works on paragraphs, full articles, social posts, emails.
✅ The bar it holds itself to
Every skill in this library self-verifies — these are this skill's own quality checks, straight from its definition.
✓Genre was assessed first (Phase 0); neutral/informational text got Phase 1 only, with no injected opinion, aside, or personality
✓Audit was completed before rewriting (patterns counted, not just detected)
✓Every removed pattern is listed in the change log with a specific reason
✓Em dashes were assessed individually — only parenthetical-substitute uses were removed
✓Rule-of-three lists: the rhythm was actually checked, not just the fact that there were three items
✓No fact, number, quote, date, or name was invented to add specificity
✓Any Phase 2 moves applied were earned by the content — none inserted to hit a quota (short sentence / "And"/"But" opener / aside added only where the material genuinely called for it, or not at all)
✓The specific detail or example added connects to an actual claim in the text, not floated in generically
⚠️ What it refuses to do
Do not fake humanity: no invented typos, no forced slang, no staged messiness, and no programmatic sentence-length variation added to hit a target — faked signals are themselves AI tells
Do not inject voice into neutral genres — documentation, summaries, reports, and official comms get the tells stripped and nothing added; personality there reads as unprofessional
Do not invent specificity — never add a number, quote, date, name, or fact that isn't real to make prose sound concrete; ask for the detail or keep the honest abstraction
Do not dismantle useful structure (a scannable layout, a genuine list) just to look less templated
Do not remove all em dashes — only the ones functioning as parenthetical substitutes should be removed; genuine dramatic pauses are valid
Install
npx pm-claude-skills add --agent claude # or codex · cursor · gemini · hermes
# or one-line MCP (every skill, any client):
claude mcp add pm-skills -- npx -y pm-claude-skills-mcp