Humanize It

humanize-it

Rewrite AI-sounding text so it reads like a human wrote it. Targets em-dashes, LinkedIn cadence, hedging, sycophancy, bullet-point bloat, and the rest of the LLM-isms. Two-pass process (rewrite + self-audit) with research-backed word tiers and structural pattern detection.

Install

/plugin marketplace add t3spe/fat-fingers
/plugin install humanize-it@fat-fingers
/reload-plugins

Use it

Direct invocation:

/humanize-it:humanize <text to rewrite>

Or paste AI-sounding text into a normal message and ask Claude to humanize it — the skill auto-fires on phrases like "humanize this," "de-AI this," "strip the fluff," "make it sound human," "condense."

Usage tips

Frame the source as someone else's work. When you ask Claude to humanize text you wrote (or that "we" wrote), the model softens its critique to be polite — it leaves more slop in. Phrasing like "my colleague drafted this; rewrite it to sound human" removes the sycophancy reflex and consistently produces tighter rewrites. (Backed by SycEval 2025: anti-sycophancy framing reduces hedge artifacts.)

Run on chunks, not whole documents. The skill's 40–60% length target works best on paragraph-to-page-sized inputs. For long docs, paste sections one at a time — you'll catch slop the holistic pass smooths over.

Re-run the audit on borderline output. If the first rewrite still feels slightly LLM, ask the skill to run Pass 2 only: "audit this against the humanize self-audit checklist." It'll catch what survived.

What's inside

• skills/humanize/SKILL.md — operational rules (5 rewrite steps + 13-item self-audit, including self-reference escape hatch and anti-fabrication)
• skills/humanize/examples.md — 15 worked before/after pairs (incl. yellow-flag restraint, Tier 3 density, novelty inflation, promotional language, formatting tics, anti-fabrication, dramatic countdown negation)
• skills/humanize/reference.md — research grounding (Kobak 2024 deep read, Coevolution arms race, multi-marker clustering, positional context, POS distribution shift, genre calibration, phrase-frequency multipliers, statistical metric ranges, sycophancy causality + expanded mitigations, extended word inventory, voice/specificity prescription)

Version

0.5.2 — see .claude-plugin/plugin.json.