| name | humanize-writing-skill |
| description | Use when text sounds "too AI" or "too ChatGPT," user asks to humanize a draft, remove em dashes, or de-AI writing. Triggers: "humanize this", "make it less AI", "sounds too ChatGPT", "remove AI patterns", "rewrite to sound human". Skip for technical docs, READMEs, code comments, commit messages, formal/academic tone.
|
| license | MIT |
Humanize Writing
Make text sound like a specific human wrote it. Read references/ai-patterns-dictionary.md first — it's the banned word list, structural tells, and examples.
Pick a voice first
- User specified → use it.
- User provided a sample → mirror it (sentence length, word choice, punctuation habits, transitions). Disliked samples are equally useful.
- Neither → ask:
"What voice?
- Clear thinker — direct, no decoration, smart person working through an idea
- Casual — like telling a friend. Warm, loose, real.
- Sharp — strong takes, punchy sentences, zero hedging
- Warm professional — credible and polished but still sounds like a person
- Your voice — paste a sample and I'll match it"
- User says "just do it" → default to clear thinker and go.
Three passes
Pass 1 — Kill AI vocabulary. Replace every Tier 1 word. Replace Tier 2 where clustered. Delete Tier 3 transition clusters (>2 in a short section) — good writing rarely needs explicit connectors. Don't just swap synonyms; restructure when the fancy word isn't needed.
Pass 2 — Break AI structures. These are stronger tells than vocabulary:
- Em dashes → commas, periods, or parentheses. Max 1 per 500 words.
- Parallel negation ("Not X, but Y") → state the positive directly.
- Rule of three → pick the one or two that matter.
- Rhetorical Q + answer ("What does this mean? It means...") → just say it.
- Mirror structures (consecutive sentences with identical shapes) → break the symmetry.
- Neat paragraph endings → let ≥30% just stop without a tidy conclusion.
- Inflation of importance ("pivotal", "testament to") → delete. Content speaks for itself.
- Signposting ("Let's dive in") → drop the announcement, start with the substance.
- Secondary convergence → when you stop using one AI pattern, don't replace it with another. Vary approaches; sometimes no transition at all.
Pass 3 — Add human texture. Apply the selected voice, then: vary sentence length aggressively (short, long, fragment); start some sentences with "And" or "But"; have opinions (remove "it could be argued"); use specific numbers/names/dates over vague claims; leave some thoughts unresolved.
Anti-AI audit
After rewriting: ask "What still makes this obviously AI-generated?" — note 2–4 remaining tells, fix them, then present the final version.
Checklist
Keep meaning and facts intact. Don't dumb it down — "human" isn't "simplistic." Don't over-correct into a different kind of artificiality — forced casualness is just as obvious. Return the rewrite only; don't explain changes unless asked.
Based on Wikipedia: Signs of AI writing, brandonwise/humanizer, blader/humanizer, lguz/humanize-writing-skill. MIT.
